How agencies track and bill AI agent work honestly

Why fake hours break trust

Mixing agent work into a person's timesheet raises utilisation for hours nobody sat, and puts a human line on machine work. Recording nothing hides effort you may be allowed to bill and loses the signal on where AI helps. The honest middle path: record agent time as the agent's next to people's time, on the same task and project, never mixed.

Three different things: human hours, agent time, and token/cost

Treat these as separate records so you can explain the cost of AI on client work without inventing hours.

• Human hours: people’s timers and timesheets; billability from task type; drives utilisation and approvals.
• Agent time: the agent logs finished duration on the task; follows task‑type billability and project rate; appears as its own row in reports (for example “Claude (agent)”).
• Tokens and model cost: usage the agent reports (model, tokens, cost) — evidence and ops data you may or may not pass through commercially.

End‑to‑end workflow: assign → task → time/cost → report → invoiced

1. Add the agent once, as a workspace member with its own token; connect it over MCP (Claude, OpenAI, Grok, Muse, Cursor, or your own script).
2. Put the work on a client task; assign the agent or @mention it in the task thread or DM.
3. Watch and steer the live session; answer questions in the thread; anyone who asked (or a manager) can stop it.
4. Optional: delegate to other agents with `agent_delegate` (max 4 hops; no loops); stopping the session stops helpers.
5. When done, the agent logs finished time on the task with `log_time` and reports model, tokens and cost.
6. Reports show the agent’s row next to people, with duration per request and usage totals.
7. Mark billable entries invoiced; they lock like anyone’s. Hourtick doesn’t generate invoices.

Agency playbook: retainers, task types and rates

Extend the skeleton you already use for people: client → project (retainer/campaign, hours budget, rate) → task types with billable flags → tasks. Agent time follows the same billability rule as people; a non‑billable project overrides everything.

• Retainers and budgets: watch burn (amber at 85%, red past 100%); agent hours on project tasks count toward the same budget view.
• Who is responsible: name both — the agent line for effort, the reviewer or account lead for quality.
• Guardrails: restrict an agent to roles, people, clients and projects; allow read‑only; set monthly budgets for usage and logged time (pauses + notifies); set expiries; keep harness instructions versioned in Hourtick. See the agent harness guide for details.

Pricing models (commercial choices, not product features)

Passing through model cost: some clients accept a usage line (tokens / provider cost), with or without markup. Either way, report usage internally so finance sees true delivery cost.

1. At the project rate (same as people): simple and transparent when agent time replaces human time one‑to‑one.
2. At a lower agent rate: honest when an agent produces less client‑facing value than a senior; treat a distinct agent rate card as a commercial overlay until product supports more rate dimensions.
3. Fixed fee per deliverable: still record agent time and usage as evidence and for margins.

What not to do

• Do not pad human hours with agent duration.
• Do not bill agent time for internal work the client never agreed to.
• Do not claim a utilisation boost from agent hours on people’s dashboards — keep utilisation about people.

Reports and invoicing: showing the work without mixing it

• Reports: filter by period, client, project, task type, person or agent; headline metrics show total hours, billable hours, billable amount and not‑yet‑invoiced; CSV export available.
• Invoice line pattern: prefer an explicit line such as “AI agent: research and drafts — 3.5 h — reviewed by [Name]”; then mark those entries invoiced so they lock.
• Utilisation stays human: agent entries never land on a person's timesheet, week total or approval.

What Hourtick does not track automatically

• Agents do not invent billable policy — task type and project settings decide billability.
• Agents do not run timers — only finished time via `log_time`.
• Token/cost requires the agent to report usage; Hourtick cannot infer provider spend.
• Hourtick does not run or resell models; provider invoices live with your AI vendor.
• Hourtick does not generate client invoices; marking entries invoiced is the bridge to your invoicing tool.
• No employee monitoring (no screenshots/apps/websites/keystrokes).
• Import from Harvest, Toggl or Clockify is not built yet.
• Delegation depth is capped (max 4 hops; no loops).
• Sessions that go quiet for 30 minutes close automatically.

What category tools miss (token dashboards vs task‑linked time)

• Task and project linkage — spend without a client task does not survive an invoice review.
• Human vs agent rows — mixing AI effort into a person's timesheet breaks utilisation and trust.
• Billability from the same task types you already use for people, with history that doesn’t rewrite after the fact.
• Invoiced locks so billed numbers can't drift when someone edits later.
• Live work context (thread, checklist, files, who asked) and visible delegation you can stop.

How Hourtick handles it (product summary)

• Agents are workspace members with their own token; connect anything that speaks MCP (Claude, OpenAI, Grok, Muse, Cursor, or your own code).
• Ask via @mention, DM or assignee; watch sessions live; answer when stuck; stop any time.
• Agents log finished time with `log_time`; it follows task‑type billability and project rate; appears as e.g. “Claude (agent)”; mark invoiced → locks; never on a person’s timesheet; agents don’t run timers.
• Reports also show duration per request and the tokens/model cost the agent reported.
• Multi‑agent: `agent_delegate` with max 4 hops; stopping a session stops helpers.
• Pricing: Free — $0 forever, unlimited people, every feature, 500 MB usage/month. Pro — $29/month for 5 GB + $10 per extra GB/month. No credit card to start. No per‑seat pricing. No feature tiers.

Get started

1. Sign up free — unlimited people, every feature, 500 MB/month, no credit card.
2. Connect agents and MCP — see Developers and AI agents.
3. Set clients, retainer projects, budgets and billable task types the way you already do for people.
4. Assign an agent to a real client task; confirm `log_time` and usage land on the task; review Reports and Invoiced tracking.
5. Deepen the harness (instructions, checks, context) — see the agent harness guide; compare plans on Pricing.

Legal note

Phrases about making AI work “billable,” putting agent lines on client invoices, or treating agent hours like human billable hours can carry contract, tax and professional‑services implications. This article describes product behaviour and common agency practice; it is not legal, tax or billing advice.