Your Copilot Is Not an Employee

It is Thursday, 3:14 PM. Your VP Operations opens the Microsoft 365 Copilot pane and types the prompt that reads "summarize this week's exec readouts and flag anything the CEO should see." Copilot ships a four-paragraph brief in 11 seconds, pulls three bullets off a Teams channel, and lands two of them against the wrong workstream. The VP reads it, edits the bullets by hand, pastes it into the Friday prep doc, and closes the pane. The brief ships to the CEO at 4:42 PM. Copilot sent zero follow-ups, logged zero decisions, and did not know the readout was due.

Your Head of Sales opens the Salesforce Agentforce sidebar and asks it to draft the renewal email for the top-10 Q4 accounts. Agentforce drafts 10 emails in a minute, pulls the CSM's last-touch note into eight of them, and misses the pricing change the pricing committee shipped on Monday because the knowledge-base article landed on Wednesday. The CSM edits six emails, kills two, and sends four. Four more sit in Outlook drafts until Monday.

Your Head of Finance opens ChatGPT with the Teams connector and asks it to reconcile the Stripe payouts against the deposits for the week. ChatGPT returns a markdown table, cites 11 of the 14 deposits, and flags a $42K variance against the wrong account. The response ends with "would you like me to open a ticket for this." Nobody opens the ticket. The variance clears on Monday when the senior accountant notices.

A chat sidebar is not an employee. The thing you bought as "the AI that works for your team" is a draft generator wired to your systems of record. It does not own a function, carry a cadence, log a decision, or close a loop. The gap is not model quality. The gap is operating shape.

What a copilot does

A copilot waits for a prompt. It reads context from a connector when the vendor shipped the integration, writes a draft back to the pane, and loses state the moment the user closes the window. Microsoft 365 Copilot at $30 per user per month, Salesforce Agentforce at $2 per conversation, Google Gemini at $20 per user per month, ChatGPT Business at $25 per user per month, and GitHub Copilot at $19 per user per month all ship on the same contract shape. Prompt in, draft out, no memory of the last five prompts past the session cache.

Pull one quarter of copilot usage across a 240-head Series C. The adoption dashboard reads 68 percent of licensed seats active in a month, 11 prompts per active seat per week, and 3 to 7 minutes saved per prompt on self-reported surveys. The CFO reads $94K a quarter in copilot license cost, 42 hours a week of self-reported savings against a loaded hour of $90 to $140, and asks which decision landed differently because of the agent. The answer is nobody kept track.

What an employee does

An employee owns a function. The Head of Finance owns month-end close. The Head of RevOps owns pipeline hygiene. The CSM owns renewal readiness for 42 accounts. The function has a cadence, a definition of done, a set of inputs, a set of outputs, and a person on the hook when the output misses. The person reads the inputs on Monday, works the queue by Wednesday, ships the output by Friday, and closes the loop on the exception by the following Monday. The cadence survives the person going on vacation because the function is written down somewhere.

A copilot ships a draft against a single prompt in a single pane. An employee ships 42 drafts, logs 11 decisions, routes 8 exceptions, follows up on 14 open items from last week, and reads Tuesday's new data against Monday's open list. The gap is three or four orders of magnitude on the volume axis and the entire ownership axis. One ships artifacts on demand. The other runs a function on calendar time.

The three gaps that make a copilot a tool

Cadence. A copilot runs when a human opens a pane and types a prompt. An employee runs on a schedule the function demands. Pipeline hygiene reviews on Monday at 9:00 AM. Close kickoff on the first business day at 8:30 AM. Renewal brief on day 90 before expiry at 10:00 AM. A copilot cannot show up at 8:30 AM on day one of close and draft the 42 workpapers off the Netsuite GL because nobody prompted it.

Memory. A copilot forgets the thread. An employee builds a working file. The CSM reads 14 renewal threads from last quarter, remembers the two accounts where legal flagged a MSA clause, and writes the pre-brief against the live account notes. A chat sidebar reads the Teams channel for this conversation and loses the thread next week. An AI employee writes to a working store with a schema the function owns, reads from it on cadence, and carries state across weeks the way a human would carry a Google Doc.

Closure. A copilot drafts an email and stops. An employee sends the email, logs the send in Salesforce, opens the follow-up in the task queue, routes the exception to the renewal lead, and reads the response when it lands. The last six steps are the function. The copilot ships step one and hands the rest back.

What a fractional AI employee runs

Hand the function definition, the inputs, the outputs, the cadence, the exception routing rules, and the connector set to a fractional AI agent. The agent runs the function. The marketing-attribution agent runs the Monday launch review, the Thursday stall digest, the Friday pipeline-reconciliation note, and the quarterly ROI brief. The close agent runs the day-one workpaper draft, the day-three variance narrative, the day-six trial balance tie-out, and the day-nine audit PBC pre-fill. The renewal agent runs the day-90 brief, the day-60 pricing confirmation, the day-30 red-flag flag, and the day-7 escalation route. The CSM reads one list on Monday instead of 42 chat windows across the week.

The agent carries state. Monday's queue reads against last Monday's open items. The exception the agent routed to legal three weeks ago reopens the moment legal replies. The quarterly ROI brief reads against the ROI brief from last quarter and names the two campaigns that trended up and the three that trended down. The chat sidebar does none of this because the chat sidebar is a chat sidebar.

The unit economics

A 240-head Series C paying $94K a quarter in copilot licenses is paying for drafts. The 42 hours a week of self-reported savings land against prompts the function would not have run anyway. The attribution team still ships a Thursday agency deck. The close still signs on day 14. The renewal brief still writes on day 7 of 90. The copilot shaved 3 minutes off 11 prompts a week per seat and changed zero function-level numbers.

A fractional AI employee for one function runs $4K to $9K a month ongoing on API spend, read-only system access, and a human operator who owns the Monday review. A 14-day sprint to stand it up runs in the low to mid five figures. Pulling close cycle from 14 days to 6, pulling marketing attribution from a Thursday agency deck to a Monday clean source table, or pulling the renewal brief from day 7 of 90 to day 90 of 90 are function-level numbers the CFO reads on the next quarterly review. The chat sidebar sits beside the function. The agent runs it.

Count the lines you want run on cadence. A Series B past $14M ARR runs 11 to 18 functions that need an owner on the hook. Three of them currently belong to a Head of X who runs the function at 68 percent and ships the other 32 to a copilot pane. Pick one, scope it, ship it in 14 days. The next quarter reads differently.

What changes after the sprint

Picture Thursday 3:14 PM, two months after the sprint ships. Your VP Operations opens the exec-readout agent inbox, reads the brief the agent drafted off the three readouts that landed Tuesday and Wednesday, confirms the two items for the CEO Friday prep, and closes the loop. The agent logs the decision, routes the follow-up to the three function leads, and queues the next readout review for the following Tuesday.

Your Head of Sales opens Monday's renewal queue and reads 14 pre-briefed accounts with the pricing change applied, the CSM note attached, and the red-flag flag raised on three. The chat sidebar still ships drafts to the AE who wants a one-off email at 4:00 PM. The function ships without the sidebar.

If your stack reads $94K a quarter in copilot licenses and zero function-level numbers moved last quarter, the version where the function runs on cadence against a working file is fourteen days away. A copilot is a draft generator. A fractional AI employee runs the function. You can keep paying for drafts, or you can scope a sprint and have the function running before the next board pack.