Shipped 01 — One Record
One checked time record — every assistant and automatic logger writes to the DNC Time Connector.
Shipped 02 — Review Engine
Merge first, then round — same-task overlaps are merged before rounding; cross-client or cross-task overlaps are flagged for review.
Shipped 03 — Three Layers
Raw record, corrections, bill — kept separate, and a person decides what to bill.
TL;DR
AI assistants now do real client work, often at the same time. If you treat each session’s clock as billable time, you will overstate the week.
We built one checked time record, the DNC Time Connector, that every assistant and logger writes to. A review step merges overlaps on the same task, flags overlaps across clients or tasks for me, and then rounds to the quarter hour. A person still decides what to bill.
We bill for work I directed and reviewed, not session time.
Why time got hard to pin down
DNC doesn’t have a large technical staff. I direct the AI assistants myself, and I keep approval authority over what they produce, including the timesheet.
Once assistants are doing client work, time stops living in one place. It’s spread across assistant logs, session transcripts, and the calendar. The weekly timesheet became something I rebuilt by hand from all of that.
What I wanted was simpler than it sounds: one time record that every assistant and every automatic logger writes to, with rules, and a review before anything is submitted.
One record, a review in the middle
The DNC Time Connector is that record. Assistants connect to it through an MCP server, and automatic loggers for sessions, meetings, and email write to the same place. Access is authenticated, and every call is audited. I review the week on a web page.
The tools the assistants get are narrow, and none of them are admin tools:
time_clients: which client, project, and task can be chargedtime_log: write a work blocktime_my_entries/time_my_hours: read back the logged time on the recordtime_retract: withdraw a block the assistant loggedtime_propose_adjustment: ask for a correction, not apply onetime_timesheet_note: write the timesheet comment for a day’s work
A logged line has a time range, a client/project/task, an optional ticket reference, a one-sentence summary of the work, and a type tag so we can tell an assistant entry from a session logger, a meeting, or an email. The type stays with the entry, which matters later when you have to explain the week.
A review engine sits between the raw log and the draft timesheet. It merges overlapping blocks for the same client and task, then rounds to the quarter hour. Work that overlaps across different clients or tasks is flagged, not merged, because deciding which one gets the hour is a judgment call. That’s mine.
Take two blocks for the same task, 9:00–9:50 and 9:30–10:20. Add them as they are and 9:30–9:50 counts twice. Round each block and you get 0.75 + 0.75 = 1.5 hours. Merge first and you get 9:00–10:20, which rounds to 1.25. The extra 15 minutes come from counting the same 20 minutes twice.
A session clock is not a workday
An AI session that’s open all day isn’t a day of work. Counting a session from its first message to its last treats it as if it had been working the whole time, and with several sessions open at once, those clocks stack up fast.
So I don’t bill from session clocks, and we’re changing the session logger to record only the stretches when a session is actually working. Tracked activity is a floor and a starting point for review. It isn’t a bill.
The bill is a separate layer
Tracked hours are not billable hours. Once work runs in parallel, that’s the whole problem.
A lot of firms are answering AI by dropping hourly billing for fixed or value pricing. For well-scoped, repeatable work, that can make sense. Oracle EBS and systems integration work usually isn’t that. You find the real problem halfway through the patch, the data fix, or the cutover. Hours are still the fairest way to share that uncertainty, as long as the hours are real. So we kept the hours and fixed the record.
There’s a second reason, and any DBA or programmer will recognize it. A long script or data load can run overnight, sometimes for days. Nobody bills for the hours it runs unwatched. You bill for the time you spend writing it, checking in on it, and dealing with what it turns up. AI works the same way. The hours measure the person, not the machine.
So the system keeps three layers and doesn’t mix them:
- The raw record. Entries aren’t edited and aren’t deleted. If an assistant retracts a block, it has 14 days to do it, and the line is marked withdrawn. It stays in the record.
- Corrections. An assistant can propose an adjustment, but it’s applied only after I approve it. An approved change is stored as a correction, not as a rewrite of history.
- The bill. I review a weekly draft for each client. I can change hours, change the comment, change the partner project, or include or leave out a line. Then I submit. Submitting locks the week. If anything was logged after the draft was built, the lock is refused and I have to review again.
Guardrails
Each assistant logs under a fixed name so its entries stay together.
Summaries are plain text, at most 300 characters, with no markup, links, or code. The timesheet line describes the work that was done. How we use AI is something we discuss with clients directly.
New client or project names go to review, and near-duplicates are refused, so we don’t grow a second catalog by typo. Entries can’t end in the future. Each entry carries a key, which stops the same backfill from being written twice.
None of this is meant to make the tracker look clever. It’s there so I can show a client the raw record, the corrections, and the bill as three different things, and explain each one.
Lawyers and tax preparers are hearing the same thing from their regulators. The ABA’s guidance on AI (Formal Opinion 512) and a 2026 IRS alert for tax practitioners come down to the same rule: bill the time you spent, not the time the tool saved. Those rules don’t govern consulting, but the principle holds. Merging before rounding, and keeping the raw record, corrections, and bill apart, is how we stay on the right side of it.
What I’d tell another consultant
- Don’t trust a session clock. First message to last message isn’t hours worked, especially when several sessions are open at once.
- Don’t bill the raw total. Measure, then decide. A person has to be able to explain the number.
- Keep one record, and make every source write to it with a type tag. If time is still scattered across chat logs and calendars, you’ll rebuild the week by hand and miss things.
- Never edit raw time. Keep a correction layer, key every backfill, and keep a name catalog, or the same client will show up three ways.
- Merge overlapping blocks before you round. Otherwise the same minutes get counted twice.
- AI is a tool. People remain responsible for the timesheet.
Questions we expect
QIf AI does the work faster, shouldn’t I pay less?
AYou pay for work I directed and reviewed, not for how long a tool ran. If a task takes less of that time, it costs less. What changes is how much gets done in a week: more tickets and projects, with better quality.
QDo you add an “AI surcharge” to the hours?
ANo. Hours are time someone actually worked. Padding them would make the timesheet indefensible, which is the whole problem the time record exists to solve.
QWho pays for the AI tools?
AWe do. AI tool and model costs are our overhead, the same as any other software we use. There’s no separate tech fee and no markup.
QHow are Attendants different?
AAn Attendant for the DNC EBS Assistant is an AI agent we offer as its own service. Attendants are subscriptions with their own fee, like any other product we sell. That’s separate from hourly work, and we never mark up hours because a tool helped.
QSo what am I actually billed for?
ADirecting the work, reviewing the code and results, testing, and explaining what changed. Reviewing what the AI produces is real, senior, billable work.
QWhy not just use the session clock?
AA session that’s open all day isn’t a day of work, and several sessions open at once would count the same hour more than once. Tracked activity is a starting point for review, not a bill.
QCan the AI change my timesheet?
AIt can’t edit or submit it. Its logged blocks flow into the draft I review. It can log a block, withdraw its own block within 14 days (the line stays in the record, marked withdrawn), or propose a correction. Only I approve corrections, and I submit the bill.
QWhat are the three layers?
AThe raw record, corrections, and the bill. The raw record isn’t edited or deleted. An assistant can propose a correction, but it’s applied only after I approve it, and the original line stays. I review the draft and submit it, which locks the week.
QCan I see the raw record?
AYes, as an extract for your work. I can show you the raw record, the corrections, and the bill as three separate things, and explain each one.
QDoes anything sensitive end up in the time record?
AIt holds what any consultant’s time record holds: work descriptions, ticket numbers and meeting details. It’s private to DNC; access is authenticated and every call is audited. Assistants follow a rule against putting personal data in summaries.
QWill my timesheet say “AI did this”?
ANo. The line describes the work done; how we use AI is something we discuss with you directly.
QWho’s responsible if something’s wrong?
AI am. AI is a tool. People remain responsible for the timesheet, and for the work.
QIf AI does it faster, won’t you have less work?
ANot so far. Getting things done faster and better surfaces the next set of issues and projects.
QIs this something I can buy?
AContact sales@davidnortonconsulting.com for inquiries.
Work With Us
If you’re wrestling with the same problem, with assistants doing real client work and a timesheet that still has to be defensible, reach out.
David Norton
David Norton Consulting
Email: sales@davidnortonconsulting.com
Phone: +1 207 741 8291
SI Field Notes is our execution log — written for prospects and partners. Outcomes and momentum, without sensitive internals.
AI is a tool. Humans remain accountable.