AI Employee Pricing Models: Credits, Seats, Tasks, or Outcomes?

§ 01What Are the Main AI Employee Pricing Models?

The main models charge for one of 5 commercial objects: access, capacity, activity, interaction, or result.

Most production quotes are hybrid.

A platform might charge a workspace subscription, 6 paid seats, pooled credits, external-model usage, storage, premium connectors, and overage. Another may charge no platform fee but require a minimum number of outcomes. A third may provide unmetered employee use for licensed people while metering autonomous or customer-facing actions.

The label alone is therefore insufficient. “Credit pricing” can behave like predictable prepaid capacity or unpredictable consumption. “Outcome pricing” can align payment with value or create disputes about what counts.

Price is a function, not a number

Write each quote as a function:

Monthly vendor cost = fixed access + seats + committed capacity + variable usage + add-ons + overage - credits/discounts

Then add the operating layers from the AI employee total-cost framework:

Total role cost = vendor cost + setup + integrations + review + correction + monitoring + expected failure cost

This prevents a clean-looking unit price from hiding the cost of producing accepted work.

§ 02How Do You Define a Billable Unit Before Comparing Prices?

Request the billing specification, rate card, or contract language for every meter.

For each unit, answer 10 questions:

1. What exact event creates the unit?
2. Is the trigger, response, action, or completed workflow counted?
3. Can one user request create several units?
4. Do retries, fallbacks, tests, errors, or abandoned work count?
5. Does model or tool choice change the multiplier?
6. When does a conversation or session close and reopen?
7. Is usage rounded up at the request, day, or billing-period level?
8. Are units pooled across roles, workspaces, or legal entities?
9. Do unused units expire, roll over, or receive a credit?
10. What happens at the limit: stop, throttle, upgrade, or overage?

Build a meter dictionary

Do not substitute your ordinary-language meaning for the contract meaning.

For example, a buyer may call a customer issue “resolved” only after 7 days without reopening. A vendor may count an outcome when the customer does not request more help within the conversation. Both definitions can be internally valid, but they are different economic units.

Map business work to billing events

Create a trace for one representative case:

Eligible case → trigger → run → model calls → retrieval → tool actions → review → retry/correction → accepted outcome

Mark every billable point. Repeat for an easy case, a difficult accepted case, an escalated case, a rejected case, and a system failure.

The trace exposes whether one “task” in the role becomes 1 task, 7 actions, 3 model calls, 2 tool fees, and 1 human review on the invoice.

§ 03When Are Subscription and Seat Models Predictable?

Subscription pricing charges for access to a plan, workspace, organization, or capacity bundle. Seat pricing charges by licensed person or user category.

They are most predictable when:

• the recurring fee is known;
• included capacity is sufficient;
• paid seat rules match real access needs;
• usage does not vary materially by role;
• overage is controlled;
• add-ons are identified; and
• unused capacity is acceptable.

Subscription strengths and risks

Calculate both utilization and effective rate:

Utilization = consumed included units ÷ purchased included units

Effective price per consumed unit = subscription cost ÷ consumed included units

A $1,000 bundle with 100,000 included units has a nominal $0.01 price per included unit. If the team uses 40,000:

$1,000 ÷ 40,000 = $0.025 per consumed unit

The nominal rate is useful for capacity. The effective rate is useful for economics.

Seat strengths and risks

Seat pricing is attractive when human access drives product value and usage is meaningfully included. It is weaker when:

• many occasional reviewers need full seats;
• different seat types have unclear permissions;
• autonomous work also consumes usage units;
• contractors or external collaborators create license complexity;
• minimum seats exceed the operating team; or
• a seat is required merely to approve an AI action.

Use:

Monthly seat cost = paid seats × monthly price per seat

Then segment paid people:

Ask whether a named user can perform more than one role, whether seats can be reassigned, and when changes take effect.

§ 04How Do Credit and Token Models Work?

Credits abstract one or more technical costs into a vendor-controlled unit. Tokens meter text or multimodal input, cached input, output, and sometimes reasoning or tool-related usage.

The benefit of abstraction is a simpler commercial currency. The risk is that buyers may not see how work becomes spend.

Credit conversion

Use:

Credits per accepted outcome = total role credits ÷ accepted outcomes

Credit cost per accepted outcome = credits per accepted outcome × effective price per credit

But first identify:

• credits per mode;
• credits per model or capability;
• credits per tool/action;
• different input and output rates;
• media, browsing, code, or computer-use multipliers;
• retries and fallback consumption;
• rounding;
• pooled usage;
• rollover/expiry;
• top-up rate; and
• balance or spend controls.

Microsoft’s current Copilot Studio billing documentation illustrates why “credit” needs a rate card: different answers, actions, grounding, flows, tools, tokens, and voice tiers can consume different quantities, and one interaction can combine several feature types.

Token economics

A simple token forecast is:

Model cost = input tokens × input rate + cached input tokens × cached rate + output tokens × output rate

An agentic workflow can add:

• orchestration calls;
• planning turns;
• retrieval context;
• tool definitions;
• tool results;
• reflection or evaluation calls;
• fallback models;
• safety classifiers;
• embeddings;
• memory reads/writes; and
• repeated attempts.

Tokens are excellent engineering signals but weak business denominators. A cheaper token does not guarantee a cheaper accepted outcome if the workflow uses more tokens, fails more often, or requires more correction.

Credit-pool scenarios

The right plan is not necessarily the bundle with the lowest nominal credit rate. It is the plan with the best effective economics and acceptable downside for the actual workload.

§ 05How Do Task, Run, Action, and Message Models Differ?

These models charge for activity, but the granularity changes.

A task is vendor-specific

Zapier’s current pay-per-task documentation describes additional task billing after plan capacity and a maximum usage limit. That behavior is specific to Zapier, but it demonstrates the procurement questions every task-based quote needs: what counts, what happens at the plan limit, who can enable overage, and where usage is visible.

An action can be more granular than a conversation

Salesforce’s current Agentforce pricing page presents several commercial units, including action-oriented Flex Credits, conversations, and user licensing. The page also shows examples in which a use case contains multiple actions.

That does not make action pricing inherently expensive or cheap. It makes workflow shape economically important.

Use:

Monthly action cost = eligible cases × AI coverage × actions per case × price per action

If accepted cases need 4 actions, rejected cases need 3, and escalated cases need 2:

Total actions = accepted cases × 4 + rejected cases × 3 + escalated cases × 2

Add retries separately. Do not assume unsuccessful work is unbilled.

Scheduled roles can create idle or duplicate activity

Persistent roles may run on a shift and schedule, react to events, poll queues, and resume unfinished work. Price:

• scheduled runs with no eligible work;
• duplicate triggers;
• polling;
• overlapping shifts;
• retries after timeouts;
• handovers;
• monitoring actions; and
• maintenance or test traffic.

One incorrectly configured trigger can change spend without creating more accepted outcomes.

§ 06Are Conversation and Outcome Models More Aligned With Value?

They can be, but only when the commercial definition matches the buyer’s definition of useful completion.

Conversation pricing

Conversation pricing can reduce sensitivity to the number of messages or actions inside one interaction. It becomes less predictable when:

• the open/close window is unclear;
• a customer returns after the window;
• channels use different rules;
• one issue spans several conversations;
• handoff opens another paid unit;
• proactive and reactive interactions differ; or
• internal and external use have different meters.

Ask for examples of:

1. one issue solved in 2 messages;
2. one issue solved in 20 messages;
3. one issue reopened after 1 hour;
4. one issue reopened after 2 days;
5. one conversation covering 2 issues;
6. one failed conversation escalated to a person.

Outcome pricing

Outcome pricing moves the billable point closer to business value. However, “outcome” remains a contract term.

Intercom’s current pricing FAQ defines several Fin outcome types and also lists cases that are not charged. That specificity is useful: buyers can compare the vendor’s billing event with their own resolution, qualification, or acceptance rule.

Use a reconciliation table:

Outcome pricing can transfer or create risk

The vendor assumes some risk when failed attempts are unbilled. The buyer may still incur:

• platform minimums;
• human review;
• correction;
• customer harm;
• tool or channel fees;
• escalation labor;
• delayed resolution; and
• integration/operations cost.

Paying only for a vendor outcome does not mean the full business cost exists only when that outcome occurs.

§ 07How Do Commitments, Overage, and Hybrid Pricing Change the Budget?

Commercial terms can matter more than the advertised rate.

Model committed, effective, and marginal rates

Committed unit rate = committed price ÷ committed units

Effective unit rate = total invoice ÷ consumed units

Marginal unit rate = cost of the next unit or usage band

All 3 matter.

A low committed rate can coexist with a high effective rate when adoption is weak. A favorable effective rate can coexist with an expensive marginal rate when the role is near its limit.

The FinOps Foundation’s FinOps for AI framework distinguishes pricing units from consumed units and highlights the need to reconcile granular AI usage - tokens, calls, and outcomes - with cost and business value.

Price the downside

For every quote, calculate:

• unused-capacity cost at 50%, 75%, and 90% utilization;
• overage cost at 110%, 125%, and 150% of forecast;
• cost if first-pass acceptance falls 10 points;
• cost if retries double;
• cost if reviewer demand rises 25%;
• cost of adding a required security or administration seat; and
• cost of exit before the contract ends.

Predictability is not the absence of variable pricing. It is the ability to explain and limit the variance before the invoice arrives.

§ 08How Do You Normalize Different Vendor Quotes?

Create one workload specification and force every quote through it.

Step 1: Define the role workload

Illustrative monthly workload:

The high case has more work and worse process efficiency. That is deliberate. A volume-only high case understates the cost of scaling.

Step 2: Translate into each meter

Covered items = eligible items × AI coverage

Normal case:

1,000 × 90% = 900 covered items

Runs = 900 × 1.15 = 1,035 runs

Actions = 1,035 × 4.0 = 4,140 actions

Then map:

Do not infer conversation or outcome count from covered tasks unless the definitions support it.

Step 3: Calculate an illustrative quote

The following rates are fictional and exist only to show normalization.

For Quote B normal:

4,140 actions × 25 credits = 103,500 credits

Included cost = $1,000

Overage = 3,500 × $0.012 = $42

The table shows $1,242 because the fictional quote also includes a $200 workspace fee. Every fixed layer must remain visible.

Step 4: Add operating cost

Assume all quotes require:

• setup amortization: $500/month;
• integration/admin: $250/month;
• review: 40 hours × $50 = $2,000;
• correction: 18 hours × $50 = $900;
• monitoring: $300; and
• expected failure cost: $250.

Non-vendor operating cost = $4,200/month

The spread between the lowest and highest full cost is much smaller than the spread between vendor prices because review and correction dominate this example.

The detailed cost-per-accepted-outcome method should be the final normalization layer, while the AI employee ROI model tests whether the accepted work creates enough attributable value.

Step 5: Reconcile the first invoice to operating events

A quote calculator predicts cost. The first production invoice tests whether the pricing interpretation was correct.

Reconcile at 3 levels:

1. Invoice: amount billed by SKU, unit, discount, tax, commitment, and overage.
2. Platform: usage recorded by workspace, role, model, tool, user, and environment.
3. Workflow: eligible cases, runs, actions, retries, reviews, and accepted outcomes.

The totals should connect:

Invoice units ≈ sum of platform usage under the contract rules

Platform usage ≈ sum of attributed workflow events plus explained shared/system usage

Small differences can come from rounding, timing, delayed records, taxes, or shared operations. Unexplained material differences need an owner and correction.

The final line is more informative than invoice variance alone. A 25% higher bill can be acceptable if accepted outcomes grew proportionally. It is a warning when accepted outcomes fall.

Separate price, usage, mix, and quality variance

Use 4 variance categories:

• Price variance: the effective unit price differs because of rate changes, discounts, overage, or commitment treatment.
• Usage variance: the same work consumes more billable units.
• Mix variance: more cases use expensive models, tools, channels, or actions.
• Quality variance: more work is rejected, corrected, reopened, or repeated.

For a simplified unit model:

Price variance = actual units × (actual unit price - budget unit price)

Usage variance = budget unit price × (actual units - budget units at actual workload)

Mix and quality need workflow segmentation rather than one blended equation. Separate easy, normal, difficult, escalation, and failed cases; then compare unit use and acceptance inside each group.

Without that separation, procurement may negotiate a lower rate while the operating team allows retries and low acceptance to erase the savings.

Reconcile non-billed operating events too

Not every economically important event appears on the invoice. Retain:

• unbilled failed runs;
• correctly escalated cases;
• reviewer time;
• correction time;
• customer or employee waiting time;
• incidents;
• unused prepaid capacity;
• plan-enforcement interruptions;
• manual work during outages; and
• tasks moved outside the platform.

These records explain why vendor spend and total role cost can move in different directions.

Set a close process

During a pilot, reconcile weekly. In stable production, complete a monthly close within 5 business days of receiving usage and billing data.

The close should produce:

Preserve the original quote model. Update a new forecast column rather than overwriting assumptions, so later reviewers can see what changed.

§ 09Which Pricing Model Is Most Predictable?

The most predictable model is the one whose meter maps cleanly to stable demand, whose failure paths are observable, and whose commercial limits are enforceable.

Score 6 dimensions:

Predictability score = sum of 6 dimensions ÷ 30 × 100

A fixed subscription may score poorly if required usage is unknown. A variable action model may score well if action counts are stable, tagged by role, and capped.

Test 5 adverse scenarios

1. Volume spike: eligible work rises 50%.
2. Quality drop: first-pass acceptance falls from 85% to 65%.
3. Retry loop: runs per item rise from 1.1 to 2.2.
4. Case-mix shift: high-complexity work doubles.
5. Noisy trigger: scheduled or event-driven activity runs 5 times more often.

For each scenario, record invoice cost, full cost, accepted outcomes, and the control that detects or stops the change.

Distinguish cost continuity from service continuity

A hard usage stop protects the budget but may interrupt customer, operations, or compliance work. Automatic overage protects service but may expose spend.

Decide:

• which roles can stop;
• which may degrade to a lower-cost mode;
• which must queue;
• which require approval for overage;
• which transfer to a person; and
• which have a protected reserve.

Budget policy is an operating policy.

Use a role-level cost dashboard

The dashboard should connect commercial meters with useful work:

Segment these metrics by role and task subtype. A blended organization average can hide one role whose long context, tools, retries, or review burden consumes disproportionate capacity.

Set thresholds before launch. For example, investigate if units per accepted outcome rise 15% week over week, pause expansion if full cost exceeds the high case, and require approval before an autonomous role crosses its monthly reserve. The percentages are examples; choose limits from observed variance and operational consequence.

§ 10What Should Procurement Ask About AI Employee Pricing?

The AI employee platform RFP checklist should request proof, not a sales explanation.

Meter and invoice

• Define every billable unit.
• Provide the complete current rate card and multiplier table.
• Show 5 representative invoice traces from trigger to charge.
• Identify rounding, minimum, commitment, and overage.
• State which failed, retried, test, sandbox, and support activities are billed.
• Show how usage is attributed to role, workspace, user, model, tool, and environment.
• Provide a machine-readable export or API.

Capacity and controls

• Show current balance and forecast.
• Demonstrate alerts at 50%, 75%, 90%, and 100%.
• Demonstrate role-level budgets.
• Explain behavior at exhaustion.
• Identify who can enable overage or change limits.
• Explain pooled-capacity priority and reserve.
• Show anomaly detection for loops or spikes.

Contract

• State renewal, price-change, and notice terms.
• Explain unused units, rollover, expiry, and refunds.
• Show true-up timing.
• Identify minimum seats and usage.
• Explain plan changes during the term.
• State data export, deletion, transition, and termination rights.
• Price support, onboarding, integration, and premium controls.

Acceptance and quality

• Define vendor outcomes separately from buyer acceptance.
• Provide dispute and correction rules.
• Explain whether human handoff is billable.
• Identify service credits and exclusions.
• Provide uptime and support commitments when relevant.

Keep a quote-comparison table:

Use the broader 12-point platform evaluation for capability, persistence, triggers, tools, permissions, logs, handoffs, data posture, and operational fit. Cheap capacity is not useful if the platform cannot support the role safely.

§ 11How Should Buyers Evaluate CellCog’s Credit Pricing?

CellCog’s public pricing page describes monthly credit bundles, additional credits, credit validity, storage, modes, and other plan features. Prices and terms can change, so verify the selected plan and checkout terms on the purchase date.

Do not preserve a price from this article in a long-lived budget. Preserve the evaluation method.

Record the live quote

Measure the role conversion

During a controlled AI employee pilot, record:

• eligible tasks;
• completed runs;
• modes used;
• tool and media activity;
• total credits;
• credits by task subtype;
• retries;
• accepted outcomes;
• review minutes;
• correction minutes; and
• idle or duplicate activity.

Calculate:

Credits per covered item = role credits ÷ covered items

Credits per accepted outcome = role credits ÷ accepted outcomes

Effective vendor cost per accepted outcome = attributed invoice cost ÷ accepted outcomes

Then add setup, integration, review, correction, monitoring, and failure exposure.

Define a CellCog budget gate

An illustrative gate might require:

• at least 100 representative items;
• first-pass acceptance of at least 80%;
• no unacceptable high-severity event;
• median review below 5 minutes;
• credits per accepted outcome within a stated band;
• no unexplained usage spike above 20%;
• normal and high cases inside budget; and
• an identified plan for unused or excess credits.

Those numbers are examples, not CellCog performance promises. Set the actual gate from the role, baseline, value, and risk tolerance.

Normalize one real role before choosing a plan: same workload, same accepted outcome, same review policy, and the same downside cases. The most useful pricing model is the one the operating team can explain, reconcile, and control.