AI Sales Agent Cost in 2026: Pricing and Full TCO

A narrow research assistant can use under $100 monthly in models. A managed outbound product may cost several thousand dollars. Voice-heavy deployments can cost much more. Data, integrations, compliance, and human review raise the total. There is no useful single market price. Vendors meter different things.

Calculate seven layers. Include platform, models, data, channels, and CRM. Add integrations, human operations, and failure recovery. Then divide by an accepted sales outcome. Prefer held qualified meetings or accepted opportunities. Do not use generated messages. The AI sales agent KPI guide defines those denominators. The broader AI sales agents guide explains where a bounded agent fits in the sales system.

I use NextLevel.AI in production for voice and multichannel agents. I use Claude Code and Codex for narrow logic. HubSpot remains the system of record. That work informs the ledger and buying rules. I did not test every AI SDR wrapper. No vendor always produces one meeting cost.

The budgets use 10,000 contacts and 2,500 monthly voice minutes. They are transparent illustrations. They are not NextLevel.AI prices or quotes. They are not verified campaign results. Replace each assumption with your own evidence.

AI sales-agent cost is a seven-layer operating model, not a subscription line, and should be reconciled against an accepted outcome with every input visibly labeled.

01 / The direct answer: AI sales-agent cost has

The direct answer: AI sales-agent cost has seven layers

The sticker price is only one line. A complete monthly total cost of ownership is:

text platform + model + data + channels + CRM/integration + people + recovery

1. Platform

This may be a seat, workspace, credit, or conversation fee. It may also be a managed retainer. Platform charges can cover the interface and orchestration. They can also cover analytics, administration, and support. Inclusions vary.

2. Model and AI usage

Some products include model use. Others pass through token charges or add a margin. Voice systems may charge separately for speech-to-text, text-to-speech, and the language model. Agent loops, retries, long prompts, retrieval, and tool calls can make the model bill larger than a one-response estimate suggests.

3. Data

An agent needs usable identity and company data. It also needs valid contact and buying-situation data. Include enrichment, validation, storage, and signal costs. Count wrong and stale records too.

4. Channels

Email needs inboxes, domains, sending tools, and monitoring. Deliverability also takes work. Voice needs numbers, carrier minutes, recording, and transcription. Branded calling can add cost. SMS and WhatsApp add message and registration fees. A multichannel fee may omit channel charges.

5. CRM and integrations

The CRM may be an existing or added cost. Include automation, custom properties, sync, and identity mapping. Add API limits, implementation, and logs. Cheap agents can create duplicate contacts. They can also overwrite clean fields. Repair can be expensive.

6. Human operations

People still define targeting and approve claims. They review exceptions and handle replies. They maintain prompts and investigate incidents. They also keep the system aligned. Count these hours at a loaded rate.

7. Failure recovery

Budget for bounced mail and blocked domains. Add poor calls and wrong-person outreach. Count duplicates, opt-out failures, and false claims. Include CRM cleanup and engineering rework. This is normal operating cost. Customer-facing models can fail.

Seven-layer total cost stack for an AI sales agent. — A useful TCO includes human work and failure recovery, not only software and tokens.

02 / Pricing models buyers will encounter

Pricing models buyers will encounter

Two tools can advertise similar monthly prices while exposing the buyer to different cost risks. Identify the billing unit before comparing the number.

Per seat or workspace

Seat pricing is predictable when a human user drives the product. It becomes less useful when agent volume, actions, credits, or channels sit outside the seat. Ask whether viewers, admins, managers, and service accounts require paid seats.

Per contact, account, or enrichment credit

This model ties spend to processed records. Ask what consumes another credit. Check repeats, refreshes, failed matches, and email checks. Check each added data source. Repeated work can raise the cost per contact.

Per action or task

An action may be one update, answer, or tool call. It can also be one longer workflow step. Salesforce uses action-based metering for Agentforce. We checked its page on August 27, 2026. It listed $500 per 100,000 Flex Credits. A standard action used 20 credits. A voice action used 30 credits. That means $0.10 and $0.15 per action. Other products and channels can add cost. Salesforce also listed $2 per conversation.

The important question is how many billable actions a successful sales workflow uses. “Research account, find contact, draft message, update record, route reply” is not necessarily one action.

Per conversation

Conversation pricing sounds outcome-aligned but usually measures product activity, not sales success. Ask when a conversation starts, when it ends, whether a returning user opens a new one, and whether failed or irrelevant conversations are billed.

Per minute for voice

Voice pricing is easier when components are explicit. We checked Telnyx pricing on August 27, 2026. The voice engine listed $0.05 per minute. That included orchestration, speech-to-text, and text-to-speech. Model use and telephony were separate. Local inbound started at $0.0032 per minute. Outbound started at $0.005 per minute. Recording and other parts can add cost.

That $0.05 is not a complete sales-agent TCO. It is infrastructure pricing. The agent still needs logic, prompts, data, numbers, integration, monitoring, QA, compliance, and a person or service that owns the workflow.

Per model token

Custom agents often pay for input and output tokens. Cached input may have another price. Tools can use other units. A token price alone gives a weak estimate. Models differ in output, retries, and task quality.

Anthropic listed Sonnet 5 at $2 per million input tokens. It listed $10 per million output tokens. We checked that in August 2026. OpenAI lists model rates on its pricing page. Its dashboard shows actual usage. These are component rates, not workflow TCO.

Managed platform or outcome fee

Some vendors charge a platform minimum or managed fee. Others tie price to meetings or pipeline. “Pay per meeting” needs a contract definition. Is it booked, held, qualified, or sales accepted? Check replacements, exclusions, attribution, and no-shows. Define active accounts too.

Custom build and implementation

A custom workflow may have low model costs. Initial labor can still be high. Count planning, integration, tests, security, launch, and maintenance. Generated code can reduce build time. It does not remove ownership.

Map of seven different billing units used in AI sales-agent pricing. — Two similar sticker prices may meter completely different units and risks.

03 / Current public pricing evidence, with labels

Current public pricing evidence, with labels

The next table separates official prices from other inputs. We checked each price on August 27, 2026. All prices may change.

Product/component	Public price or model	Evidence label	What the price does not prove
Telnyx Voice AI	$0.05/min voice engine; LLM and telephony extra; outbound carrier rate listed from $0.005/min	Official current price	Complete agent, data, integration, QA, or sales result
Salesforce Agentforce	$500/100,000 Flex Credits; 20 credits per standard action, 30 per voice action; $2/conversation option	Official current price	Number of actions per sales outcome or required Salesforce TCO
HubSpot Credits	$0.010 per additional credit; 500/3,000/5,000 monthly credits listed for Starter/Professional/Enterprise Sales Hub	Official current price	Credits consumed by a specific agent workflow or base seat cost
Anthropic Claude Sonnet 5 API	$2/million input tokens; $10/million output tokens	Official current price	Total orchestration, tools, retries, hosting, data, or QA
NextLevel.AI	No public price used in this model	Not publicly verified here	Requires a current scope and quote
AI SDR wrapper market ranges	No number treated as verified	Third-party or author-supplied estimate only	Actual contract, limits, service scope, and outcome definition

HubSpot’s official Sales Hub pricing page states that paid tiers include monthly credits and that additional credits cost $0.010 each. It also says unused credits reset monthly. The base product and seat configuration still matter. A credit forecast that excludes the CRM subscription is an incremental estimate, not full TCO.

The author supplied market ranges for wrappers and meetings. No current quote or sample was provided. We exclude those figures from the factual table.

Four labels for official prices, current quotes, observed costs, and model assumptions. — Every number in a useful pricing model needs a provenance label.

04 / The complete TCO ledger

The complete TCO ledger

Build the ledger before requesting vendor demos. It forces every option into the same boundary.

Cost line	Billing unit	Evidence to collect	Common omission
Agent platform	seat, workspace, minimum, action, conversation	quote, rate card, invoice	annual commitment and overage
LLM/model	input/output tokens, request, minute	usage export, model rate	retries and tool loops
Speech	STT/TTS minute or character	rate card, bill	premium voices and language mix
Telephony	minute, number, recording, registration	carrier invoice	outbound versus inbound rate
Email	inbox, domain, message, warming/monitoring	sending bill	replacement domains and support
Data/enrichment	record, credit, lookup, refresh	provider invoice	repeat and failed lookups
CRM	seat, tier, credit, API	contract and usage report	mandatory base subscription
Integration	setup hour, connector, run	project log	changes after launch
Monitoring	logs, alerts, traces, storage	infrastructure bill	retention and incident review
Human QA	hour or reviewed interaction	time log	manager and compliance time
Rework	incident, cleanup hour, lost channel	incident log	opportunity cost and reputation

For every line, record one of four labels:

Official price: visible on a current primary source.
Current quote: supplied for the specific company and scope.
Observed cost: appears on an invoice or usage export for a defined period.
Model assumption: chosen for planning and not claimed as a market fact.

Do not blend the labels into one confident-looking total. A useful model can contain assumptions, provided the reader can replace them.

05 / Three illustrative monthly budget scenarios

Three illustrative monthly budget scenarios

These scenarios use the owner-supplied workload of 10,000 contacts and, where relevant, 2,500 voice minutes. They are planning examples. They do not represent vendor quotes or campaign performance.

Scenario A: research-and-draft copilot

The agent researches records, drafts possible messages, and writes structured notes for human approval. It does not autonomously send.

Line	Illustrative assumption
Enrichment and validation	$600
Model and retrieval usage	$120
Orchestration, hosting, and logs	$180
CRM incremental cost	$0, assuming existing HubSpot capacity
Human review: 20 hours at $45	$900
Maintenance: 6 hours at $65	$390
Failure/rework reserve	$160
Illustrative monthly TCO	$2,350

The model cost is not the dominant line. Human review and data are. If better rules reduce irrelevant records, the workflow may save more through fewer reviews than through a cheaper model.

If the team approves 1,000 researched records, the illustrative cost is $2.35 per approved record. That is not a revenue outcome. Continue the denominator through replies, held meetings, and accepted opportunities.

Scenario B: voice and multichannel agent

The agent processes 10,000 contacts and uses 2,500 outbound voice minutes, with human review for exceptions and replies.

Line	Illustrative assumption
Voice engine: 2,500 × $0.05	$125.00 official component arithmetic
Outbound telephony: 2,500 × $0.005	$12.50 official starting-rate arithmetic
LLM, numbers, recording, messaging	$110 assumption
Enrichment and validation	$650 assumption
Agent/orchestration operating layer	$700 assumption; not a NextLevel.AI price
CRM incremental cost	$0, assuming existing capacity
Human QA and reply handling	$900 assumption
Monitoring, compliance, rework	$450 assumption
Illustrative monthly TCO	$2,947.50

At 50 held qualified meetings, that model yields $58.95 per meeting. At 75 it yields $39.30. At 100 it yields $29.48. The system did not become cheaper; the denominator changed. This is why a cost-per-meeting claim without the meeting definition and count is incomplete.

The author supplied a $28–$35 meeting-cost range. It covered NextLevel.AI plus custom orchestration. The full ledger and period were absent. We also lack the held and qualified counts. We cannot verify that result. The $29.48 model point is illustrative. It does not confirm the supplied range.

Scenario C: CRM-native action-based agent

This scenario keeps the workflow close to the system of record and meters actions or credits.

Line	Illustrative assumption
12,000 standard actions × $0.10	$1,200 official Agentforce-rate arithmetic
Base CRM and eligible edition	Excluded; obtain company-specific contract cost
Data and enrichment	$600 assumption
Admin, flow changes, testing	$750 assumption
Human QA and exception handling	$700 assumption
Failure/rework reserve	$250 assumption
Illustrative incremental TCO	$3,500 plus base CRM

This is not a Salesforce quote. A real workflow could use fewer or more actions, and eligible editions, licenses, integrations, voice, or data may change the total. The value of a CRM-native agent is not automatically lower unit cost. It may be cleaner governance, identity, permissions, and write-back.

HubSpot Credits require the same discipline. Estimate which features consume credits. Subtract the included monthly allotment. Add extra credits at the current rate. Keep the base Sales Hub price visible.

Three illustrative AI sales-agent budget compositions with official components, assumptions, and excluded base costs. — The dominant cost can move from model usage to data, people, or the underlying CRM.

06 / How to calculate cost per accepted outcome

How to calculate cost per accepted outcome

Select a denominator that finance and sales agree on before the pilot.

Cost per delivered activity

TCO / delivered messages or completed calls

Useful for infrastructure monitoring, but weak for a buying decision. High activity can coexist with poor targeting.

Cost per positive reply or qualified conversation

TCO / contacts meeting the defined reply or conversation rule

Define positive reply, unique contact, and qualification. Exclude auto-replies and support requests.

Cost per held qualified meeting

TCO / meetings that occurred and passed the qualification rule

This is stronger than booked meeting because it accounts for no-shows and bad bookings. Record whether sales accepted the meeting.

Cost per accepted opportunity

TCO / opportunities created under a written CRM rule and accepted by the owner

Use this measure when volume and cycle length allow it. Keep the credit window visible. Exclude prior opportunities under the written rule.

A denominator ladder

For each month, show the same total cost divided by:

contacts processed;
delivered touches;
qualified conversations;
held qualified meetings;
accepted opportunities.

This exposes where cost or quality deteriorates. A vendor may look inexpensive at the activity layer and expensive at the opportunity layer.

07 / Sensitivity matters more than one forecast

Sensitivity matters more than one forecast

Use at least four tests.

Outcome sensitivity

If the illustrative voice/multichannel TCO is $2,947.50, cost per held meeting is:

Held qualified meetings	Cost per held meeting
25	$117.90
50	$58.95
75	$39.30
100	$29.48

Human-review sensitivity

If low-confidence outputs double, QA hours may dominate the budget. Model 5%, 15%, and 30% exception rates using the real minutes per review.

Data-quality sensitivity

Model the cost when 10%, 25%, or 40% of records are unusable. Include wasted enrichment, message or call charges, review time, and suppression work.

Action-count sensitivity

For credit-based platforms, estimate the workflow at a minimum, expected, and failure-retry action count. One contact can create multiple billable actions.

Illustrative chart showing cost per held meeting falling as held qualified meetings rise at a fixed modeled TCO. — A cost-per-meeting claim can move dramatically when the outcome count changes.

08 / Hidden costs that change the result

Hidden costs that change the result

Deliverability and channel reputation

Email costs extend past each message. Teams must set up and protect domains. They must manage inbox health and access checks. Monitoring and recovery also take time. Domain damage can stop several campaigns.

Identity and suppression

Every channel must honor opt-outs. It must protect customers and open deals. It must also follow call limits. Duplicate identities need a clear rule. One bad merge can trigger cleanup and unwanted contact.

Prompt and policy maintenance

Offers and prices change. Competitors and product claims change too. Security positions and territories also move. Someone must update and test approved knowledge. A prompt is not permanent.

Observability

Store the source record, prompt or policy version, tool calls, output, send decision, CRM write-back, and error. Without traces, a team cannot explain spend or investigate failures.

Human takeover

Sensitive replies need fast human ownership. This includes pricing, legal, security, and complaints. Unusual buyer context may also need help. Measure takeover delay and repeated handoffs.

Compliance and consent

Calls and messages can require legal controls. So can retention and cross-border data. This guide is not legal advice. Use qualified counsel when required.

Implementation and change management

The first version is not production cost. Add security review, tests, and data mapping. Add training, rollback, and monitoring. Budget for CRM and API changes.

09 / Build versus buy: the boundary I use

Build versus buy: the boundary I use

I would build narrow logic when the company owns the rule. Good examples include routing and enrichment order. Research structure and exception logic also fit. CRM write-back can fit too. Claude Code and Codex can speed this work.

I would buy hard infrastructure. Reliability, permissions, and compliance are costly to reproduce:

CRM and system of record;
telephony and number management;
email delivery infrastructure;
identity and enrichment data;
consent, retention, and audit controls;
monitoring or security components with mature operational requirements.

Do not replace every SaaS product. SMB teams can unbundle selectively. Keep durable infrastructure. Remove a thin wrapper only when ownership is cheaper. Keep a person responsible for each external action.

Build when

the workflow is narrow and stable;
the team can specify the rule and exceptions;
required APIs and data rights exist;
a named owner will monitor and maintain it;
the failure impact is bounded and reversible.

Buy when

carrier, security, identity, compliance, or uptime is core;
multiple regions and channels require continuous operations;
formal permissions, audit, support, and contractual commitments matter;
internal maintenance would become another product team;
the vendor’s operating depth is more valuable than interface customization.

Combine when

the CRM remains authoritative;
a telecom or sending vendor provides the channel;
a model provider handles inference;
custom logic makes the sales decision explicit;
humans approve risky or ambiguous outcomes.

Build-versus-buy architecture separating narrow custom logic from durable sales infrastructure. — Generated code can replace a thin workflow layer; it does not remove infrastructure or ownership.

10 / A 30-day pilot budget

A 30-day pilot budget

Do not start with an annual cost-per-meeting promise. Start with a capped pilot and evidence requirements.

Week 1: define scope and baseline

Choose one segment, one offer, one channel sequence, and one qualification rule. Record the current human workflow and its loaded cost. Exclude accounts with active opportunities or suppression rules.

Week 2: run in draft or shadow mode

Let the agent research, recommend, or simulate without autonomous sends. Measure data failures, unsupported claims, duplicate identities, action counts, model use, and reviewer minutes.

Week 3: controlled production

Release only low-risk cases under volume and spend caps. Keep human approval for uncertain claims and sensitive replies. Verify CRM writes and suppression behavior daily.

Week 4: reconcile invoices and outcomes

Collect every vendor invoice. Add model, data, channel, and labor records. Add incidents, meetings, and opportunities. Reconcile these records to actual spend. Prefer held or qualified meetings over bookings.

The pilot should have a stop condition for spend, negative feedback, opt-out errors, duplicate outreach, or CRM corruption.

11 / A copyable AI sales-agent cost worksheet

A copyable AI sales-agent cost worksheet

The worksheet below is designed for procurement and pilot review. It forces one owner to classify each number. It also exposes missing evidence before a contract is signed.

Workload definition

Name the target market.
Name the target segment.
Count unique accounts.
Count unique contacts.
Count planned messages.
Count planned calls.
Estimate voice minutes.
Count phone numbers.
Count sending domains.
Count active inboxes.
Count supported languages.
Count CRM users.
Count required integrations.
Define the pilot period.
Define the production period.

Do not use annual volume alone. Monthly peaks can change concurrency and support needs. Regional mix can change carrier rates.

Platform evidence

Save the public rate card.
Save the checked date.
Save the current quote.
Name the billing unit.
Record the minimum commitment.
Record the contract term.
Record included usage.
Record overage rates.
Record credit expiry.
Record cancellation terms.
Record support level.
Record implementation fees.
Record required seats.
Record administrator seats.
Record taxes separately.

A sales deck is not a rate card. A verbal promise is not a contract term. Keep the source beside the number.

Model and agent usage

Name the model.
Save the model version.
Record input tokens.
Record cached input.
Record output tokens.
Record tool calls.
Record failed calls.
Record retry attempts.
Record retrieval operations.
Record average prompt size.
Record average response size.
Record peak request rate.
Record agent loop limits.
Record safety-model use.
Record model switching rules.

Use the provider usage export when possible. A prompt estimate can miss retries. It can also miss hidden tool work.

Data and identity

Name every data provider.
Record each credit unit.
Record lookup volume.
Record refresh volume.
Record failed lookups.
Record duplicate records.
Record invalid emails.
Record invalid numbers.
Record stale job titles.
Record missing company fields.
Record suppression matches.
Record existing customers.
Record active opportunities.
Record enrichment latency.
Record data cleanup hours.

Bad data creates several costs. It wastes credits. It also wastes review time and buyer trust.

Keep the channel invoice separate from the agent platform. This makes a provider change easier to model.

CRM and integration costs

Record the base CRM tier.
Record paid seats.
Record included credits.
Record added credits.
Record API limits.
Record connector fees.
Record automation runs.
Record custom objects.
Record custom properties.
Record implementation hours.
Record test hours.
Record failed syncs.
Record duplicate creation.
Record cleanup hours.
Record change requests.

Mark existing CRM cost as base cost. Mark added capacity as incremental cost. Show both in the final model.

Human operations

Name the workflow owner.
Name the sales owner.
Name the data owner.
Name the compliance owner.
Record targeting hours.
Record prompt hours.
Record review hours.
Record reply-handling hours.
Record manager hours.
Record support hours.
Record incident hours.
Record training hours.
Record loaded hourly rates.
Record weekend coverage.
Record takeover latency.

Do not value human time at zero. Existing employees still spend finite hours. Those hours have another possible use.

Failure and rework

Count wrong-person contacts.
Count duplicate touches.
Count unsupported claims.
Count opt-out failures.
Count suppression leaks.
Count bad CRM writes.
Count missed replies.
Count dropped calls.
Count poor recordings.
Count blocked domains.
Count buyer complaints.
Count manual corrections.
Count rollback events.
Record recovery hours.
Record direct recovery cost.

Some failures cost money at once. Others create future risk. Track both without inventing a dollar value.

Outcome data dictionary

Define a delivered message.
Define a completed call.
Define a unique contact.
Define a human reply.
Define a positive reply.
Define a qualified conversation.
Define a booked meeting.
Define a held meeting.
Define a qualified meeting.
Define a sales-accepted meeting.
Define an accepted opportunity.
Define an attributed opportunity.
Define a no-show.
Define a reschedule.
Define a disqualified account.
Define the attribution window.
Define the opportunity owner.
Define the CRM creation rule.
Define the deduplication rule.
Define the reporting timezone.

Write these definitions before launch. Do not repair them after results appear. Stable rules reduce selective reporting.

Monthly reconciliation

Export each vendor invoice.
Export model usage.
Export channel usage.
Export enrichment usage.
Export CRM credits.
Export agent logs.
Export review hours.
Export incident records.
Export held meetings.
Export accepted opportunities.
Match invoice dates.
Match usage timezones.
Remove duplicate outcomes.
Separate base costs.
Separate one-time costs.
Label every assumption.
Explain every variance.
Save the final model.
Name the approver.
Set the next review date.

The reconciliation creates the evidence behind a cost claim. Without it, a precise number is still an estimate.

12 / Example: reconciling one month without hiding the

Example: reconciling one month without hiding the base costs

Assume the pilot ends on August 31. Finance exports every invoice for August. RevOps exports the same period from HubSpot. Engineering exports model and agent usage in UTC. The sales report uses the company reporting timezone.

The team first aligns the dates. A late carrier invoice belongs to August usage. It should not move into September analysis. A prepaid annual platform fee needs monthly allocation. One-time implementation stays visible in a separate column.

Next, the owner separates base and incremental costs. HubSpot may already support the sales team. Its base subscription remains part of full TCO. Added credits belong in the incremental view. Both views answer different budget questions.

The owner then checks voice minutes. Agent-runtime minutes may differ from carrier minutes. Recording can use another minute count. Failed calls may still create carrier or lookup charges. Each invoice needs its own unit.

Model usage requires the same care. Input and output tokens have different rates. Cached input can have another rate. Retries can create cost without producing another buyer contact. The usage export should explain each total.

Data credits also need reconciliation. Ten thousand contacts do not guarantee ten thousand lookups. Some records need several providers. Some fail validation. Some are suppressed after enrichment. Those events still affect cost.

The human ledger comes next. Reviewers record actual minutes, not rough memory. Reply handling stays separate from model monitoring. Incident work stays separate from routine maintenance. Loaded rates should match the finance method.

Now the team cleans the outcome report. Booked and held meetings stay separate. Duplicate meetings count once. Existing opportunities leave the attributed set. Sales applies the written qualification rule before seeing cost per outcome.

Suppose the model shows 75 held qualified meetings. The illustrative TCO is $2,947.50. That creates a modeled $39.30 per held meeting. The arithmetic is simple. The evidence behind both numbers is the hard part.

The owner then checks opportunity acceptance. Perhaps sales accepts 18 opportunities. The same modeled TCO becomes $163.75 per accepted opportunity. This denominator is smaller and commercially stronger. It may change the buying decision.

One month cannot prove a durable return. Account mix may change next month. Sales capacity can also change. Seasonality, offer quality, and follow-up speed may affect outcomes. Keep these limits beside the result.

The final review compares forecast and actual cost. Every material variance gets an explanation. More retries may raise model usage. Poor data may raise review time. Slow takeover may reduce accepted outcomes.

Currency also needs a fixed rule. Choose one reporting currency for the model. Save each source currency beside its rate. Record the exchange date and source. Do not mix converted and unconverted invoices.

Keep tax treatment in a separate field. Some prices exclude sales tax or VAT. Some carrier fees include local surcharges. A tax-inclusive invoice cannot match a tax-exclusive rate card without adjustment.

Discounts need their own record. Note the standard price first. Then record the discount, term, and expiry. A first-year discount should not define the renewal forecast. Model the next contract year too.

Prepaid credits also affect cash flow. Expense recognition may differ from cash paid. Record both views when finance needs them. Unused credits can turn a low unit price into waste.

Finally, record who can change each assumption. Procurement owns contract terms. Finance owns loaded rates and tax rules. RevOps owns sales definitions. Engineering owns usage and failure data. One worksheet can still have several accountable owners.

This reconciliation creates three outputs. Finance receives the complete ledger. Sales receives the outcome funnel. Engineering receives the failure and usage report. All three views use the same period.

The team can now set a renewal rule. It can cap spend per accepted opportunity. It can cap error and suppression failures. It can also require lower review time. Those rules are stronger than a vendor activity promise.

Repeat this process every month during the pilot. Preserve the raw exports and formulas. Change assumptions only with a dated note. That creates an honest cost history.

13 / Vendor pricing questions

Vendor pricing questions

What exactly is the billing unit?
Which actions, retries, failed calls, or returning conversations are billable?
Which model, speech, telephony, data, and channel costs are included?
Is there a platform minimum, annual commitment, prepaid credit, or expiration?
Which users require seats?
What are the setup, implementation, migration, and support fees?
What volume, concurrency, API, storage, and retention limits apply?
Can we export raw usage and map it to an interaction or CRM record?
How are opt-outs, duplicates, active opportunities, and customer exclusions enforced?
What happens when the agent or integration fails?
Which pricing is current and generally available, rather than a promotional or roadmap offer?
How will the contract define a meeting or outcome if pricing depends on it?

14 / FAQ

FAQ

How much does an AI sales agent cost per month?

A narrow assistant may use under $100 monthly in models. Production workflows can cost thousands. Platforms, data, channels, CRM, people, and rework raise cost. Use a complete ledger, not one vendor price.

How much does an AI voice sales agent cost?

Voice infrastructure can use minute pricing. Telnyx listed a $0.05 voice engine rate. We checked it on August 27, 2026. Model and carrier fees were separate. A full agent also needs logic and data. Add integration, review, controls, and human takeover.

Are AI SDRs cheaper than human SDRs?

They may cost less for narrow sales work. Research, drafting, routing, and support can fit. A fair test needs the same market and workload. It also needs one quality and outcome rule. This packet cannot support a universal benchmark.

Is a custom AI sales agent cheaper than SaaS?

It can be cheaper for a narrow workflow. The company must own its upkeep. Missing integration, security, monitoring, and support raise cost. Generated code reduces build effort. It does not remove operations.

What costs sit outside an AI sales-agent subscription?

Common external costs include models, data, channels, and CRM. Add connectors, setup, monitoring, and manager review. Add controls, incident work, and data cleanup.

What cost metric should a sales team monitor?

Track monthly TCO per held qualified meeting. Also track cost per accepted opportunity. Show contacts, actions, and conversations underneath. Keep booked and held meetings separate.

How often should pricing be rechecked?

Recheck before purchase and renewal. Recheck after major workflow changes. Check usage rates each quarter. Store each date and source beside its rate.

Research note

Methodology

01The guide combines current official pricing pages, an explicit seven-layer ledger and illustrative budget scenarios.
02Illustrative scenarios and sensitivity models are decision aids, not vendor quotes or market averages.
03Mutable prices, limits and packaging were reviewed on 27 August 2026 and must be rechecked before purchase.

Read the full methodology

Source ledger

Sources & editorial notes

01
Telnyx Voice AI Agent Pricing
telnyx.com · official pricing page; reviewed 2026-08-27. Strong usage-cost component; telephony/model/add-ons and implementation still need full ledger.
02
Salesforce Agentforce Pricing
salesforce.com · official pricing page; reviewed 2026-08-27. Current CRM-agent pricing evidence; consumption rules and total ecosystem cost matter.
03
HubSpot Sales Hub pricing and credits
hubspot.com · cited source; reviewed 2026-08-27. Recheck mutable scope, pricing and availability before implementation.
04
Anthropic Claude Sonnet 5 pricing announcement
anthropic.com · cited source; reviewed 2026-08-27. Recheck mutable scope, pricing and availability before implementation.
05
OpenAI API pricing
openai.com · cited source; reviewed 2026-08-27. Recheck mutable scope, pricing and availability before implementation.

Corrections or primary material: contact the corrections desk.