This guide is for the operations lead who owns a live agent and the finance owner who defends its budget; the agentic AI automation service covers delivery. Runtime traces, budget ceilings and kill switches stop spend inside a run; this page measures what each finished outcome cost. Reviewer capacity planning sizes the review queue, and measuring AI automation ROI after launch tracks the benefit side.
In this guide
- Why is the model bill the wrong number for an agent's cost?
- What counts as an accepted task?
- The cost-per-accepted-task ledger
- Setting up the measurement
- Keep, narrow, re-route or retire: reading the ledger
- When not to measure yet
- Where AI fits
- What delivery record exists, and what does not?
- Alternatives and selection criteria
- What are the limits of this guide?
- Frequently asked questions
- Next step
Why is the model bill the wrong number for an agent's cost?
The model bill counts tokens, not outcomes. It is keyed to an API key and a model name, so it cannot say which workflow, which task type or which retry produced the spend; Datadog's write-up on agent cost makes the same point when it ties estimated cost to the traces and applications that produced it rather than to the key.
The larger costs of a live agent usually sit outside the bill. Every escalation to a person, every draft a reviewer corrects and every case reopened after the agent closed it costs staff time. Research comparing agents with human workers found agents far cheaper and faster per task, but also found lower-quality output that agents sometimes masked, which is exactly the work a human reviewer then absorbs. Cost per run hides that; cost per accepted task shows it.
What counts as an accepted task?
An accepted task is one that met the acceptance rule agreed before measurement started: the refund was correct, the ticket stayed closed, the extracted fields matched the source document. Write it down per task type before the first measured week and keep it stable.
Deflection is not acceptance. A conversation that never reached a person may still have left the customer's problem unsolved, so count task completion separately from containment and only count a task once the outcome holds. A reopen within a short window moves the task back out of the accepted column.
The cost-per-accepted-task ledger
Five lines make up the ledger. Each needs a field in the trace or the ticket, otherwise the line becomes a guess.
| Ledger line | What it includes | Trace or ticket field it needs |
|---|---|---|
| Run spend | Model tokens, tool and API calls, retrieval, for every run in the period, including failed and abandoned runs | Task id, task type, per-step token and tool cost, outcome label |
| Hand-off time | Minutes a person spends taking over a task the agent escalated, including reading the context it handed over | Hand-off reason, hand-off timestamp, minutes until closed by a person |
| Review and rework time | Minutes spent approving, correcting or redoing agent output before it is accepted, and after a reopen | Reviewer id, review minutes, correction flag, reopen flag |
| Maintenance share | Prompt and tool updates, evaluation runs, incident time and monitoring for the agent, spread over the period | Change log entries and engineering time booked to the agent |
| Accepted tasks | The denominator: tasks that met the acceptance rule and did not reopen | Outcome label against the written acceptance rule |
Cost per accepted task is the sum of the first four lines divided by the fifth. Price the human minutes with your own fully loaded labour cost for that role; the ledger needs your figure, not a market average. Failed and abandoned runs stay in the numerator on purpose: an agent that fails half its runs costs roughly twice its per-run spend for every success.
Hand-off reason is the field most teams forget. Without it, escalations are one number; with it, you can see whether the agent hands off because a policy says so (often correct and cheap), because confidence was low (a tuning problem), or because it misread the request (an evaluation gap).
Setting up the measurement
-
Write the acceptance rule per task type
Name what "done correctly" means and the reopen window, before any data is collected.
-
Tag every trace with a task id and a task type
Spend that cannot be joined to a task cannot be attributed; this is the same per-step trace the runtime controls already rely on.
-
Log hand-off reason and human minutes in the ticket or approval tool
Use a short fixed list of reasons, so the categories can be counted, and record minutes when the person closes the task.
-
Measure the human baseline on the same task types
Record how long people take per task, how often their work needs correction, and the volume per month, for the same acceptance rule.
-
Run one full stable period before deciding anything
-
Review the ledger with the agent owner and finance together
Read the cost per accepted task next to the acceptance rate and the hand-off reasons, then pick a decision from the matrix below.
Keep, narrow, re-route or retire: reading the ledger
-
Cost per accepted task well below the human baseline, acceptance stable, hand-offs mostly policy-driven
- Decision
- Keep and widen carefully
- Why
- The agent works on this task type; widen one task type at a time and re-measure
-
Low cost on some task types, high rework on others
- Decision
- Narrow
- Why
- Remove the task types where review and rework cost more than a person doing the work
-
Most hand-offs are low confidence or misread requests
- Decision
- Re-route and re-evaluate
- Why
- Fix routing, retrieval or prompts and re-run the evaluation set before trusting the numbers again
-
Cost per accepted task at or above the baseline after a stable period, with no fixable cause
- Decision
- Retire or return the work to people
- Why
- A cheap model bill does not rescue an expensive review queue
When not to measure yet
If volume is so low that one bad week swings the result, or there is no stable acceptance rule or scope yet, per-task cost gives false precision; run the pilot on quality first. Measurement also costs time, and a small internal agent may justify only a quarterly check. For the loop a vendor runs on its own AI-run products, see the lessons guide.
Where AI fits
Netbase works with the major commercial and open-source AI tools and models, chosen per project, and treats task tagging, hand-off logging and an acceptance rule as part of making an agent fit to operate.
- In delivered work
WhatsApp AI chatbot with CRM integration
An AI chatbot for WhatsApp Business synchronised with the client's CRM, delivered in milestones from design and prototype through integration, for a client that is not named; a CRM-connected agent is where hand-off reasons can be logged per conversation.
- In delivered work
AI content moderation on a classifieds marketplace
AI flagging delivered with an admin review dashboard, for a client that is not named; a human review step like this is the review-and-rework line of the ledger.
- In delivered work
4over4's recommendation engine
It suggests products from browsing history and previous purchases.
What delivery record exists, and what does not?
- What exists. Netbase designs, builds, integrates and runs software across technologies and industries. Beyond the 4over4 recommendation engine, Netbase has delivered AI projects for clients that are not named, including retrieval-augmented knowledge assistants, document AI and MLOps pipelines, plus the chatbot and moderation records above. Most Netbase projects are delivered on fixed-price contracts, with milestone-based, monthly team retainer and KPI-linked models also offered.
- What does not. No published Netbase record states a cost per accepted task, a token spend, a hand-off rate, a saving or a payback period for any agent. The ledger above is a method, not a reported result. For co-build or build-operate arrangements, the terms, including who carries an agent's running cost, are agreed deal by deal with the CEO and never published; see the engagement models and build-operate-transfer for software products.
Alternatives and selection criteria
| Approach | Strength | Weakness | Choose it when |
|---|---|---|---|
| Model bill only | No setup; available from the provider today | Ignores human time and failed runs; cannot attribute spend to a task | An early, internal pilot with low volume |
| Cost per run from traces | Attributes spend to workflow and task type | Still excludes hand-off, review and rework time | Engineering wants to find expensive steps or loops |
| Cost per accepted task (the ledger) | One number that includes human time and failures, comparable with the human baseline | Needs an acceptance rule, ticket fields and a baseline | The agent is live, has stable volume and someone must decide whether it stays |
Three questions pick the row: is the agent live with stable volume, does a person touch a meaningful share of its tasks, and does anyone need to decide whether it stays. Two yes answers point to the ledger; cost per resolution only means something next to quality signals such as reopen rate.
What are the limits of this guide?
- This is a measurement method, not an audit or a financial guarantee; the ledger lines must be adapted to the tools, tickets and labour costs of the company running the agent.
- Published practice on agent unit cost is young, so expect to revise acceptance rules and hand-off categories after the first periods.
Plan the next step with a Netbase consultant
Frequently asked questions
Cost per run divides model and tool spend by the number of runs. Cost per accepted task adds human hand-off, review and rework time and a maintenance share, counts failed runs, and divides only by the tasks that met the acceptance rule, so it is the number to compare with a person doing the same work.
Yes. Review, correction and hand-off minutes are work the agent creates, and leaving them out makes an agent look cheaper the more a person has to fix its output. Price those minutes with your own fully loaded labour cost for the role.
Measure at least one stable period after setup, with a written acceptance rule and the same rule applied to the human baseline.
Next step
Tell us which agent is live, or about to be, and which task types it handles, and we will book a solution review to set up its acceptance rules, hand-off reasons and cost-per-accepted-task ledger. You can also see AI automation and agents or the enterprise AI agent platform.
Related services and solutions
AI automation and agents that keep people in charge Netbase provides AI automation and agent development for operations teams that want repetitive, multi-step work done by software while people keep approval over decisions that matter. We combine rule-based automation with AI where it adds value, build in human approval, and measure return against a baseline.
Learn More
AI customer service automation connected to your orders and CRM, with human hand-off AI customer service automation is a support layer where AI chat answers routine questions and triages tickets using live order and CRM data, then hands anything sensitive or uncertain to a person with the full context. Netbase builds it as a growth capability, starting from its AI chatbot and WorkChat integrator module.
Learn More
Enterprise AI agent platform: run AI agents with approvals, audit logs and cost limits An enterprise AI agent platform is a governed runtime where AI agents plan and carry out multi-step work through approved tools, while approvals, audit logs, access rules and cost limits keep every action accountable. Netbase offers it as a forward-looking solution, built inside your own environment rather than licensed as a product.
Learn More
Work with Netbase remotely
Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.
Discuss a project
Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.project@netbasejsc.com
+84 937 869 689
91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam
Get in touch
Tell us what you want to build, modernize, or operate.