The delegation ladder: what to hand over to AI first
· Prabhu Eshwarla
In my last piece, I argued that the future of business software is governed delegation: AI doing the work between our systems, with policy deciding what it may do, deterministic systems executing it, and evidence of everything that happened.
That raises a practical question. If you accept the idea, where do you start? What would you actually hand over first?
The answer is the same as with a new pilot. No airline puts a new pilot in command on day one, however talented. They fly as first officer, log their hours, and earn the captain's seat on their record. Delegating to AI works the same way.
I think of it as a ladder with seven rungs. On the lower rungs, AI reads and checks while people decide. Higher up, AI acts on its own within limits the organisation sets.
The seven rungs
1. Finding answers
"Which contracts renew this quarter?" "What is our notice period for senior managers?" "What did we decide about pricing in March?" AI answers from your own documents and cites the source, so anyone can check it in seconds. Nothing changes in the world; the risk is a wrong answer, and the citation makes that easy to catch.
2. Reading and extracting
Invoice details, contract terms, application fields, pulled from documents into a structured table. This is work people do by retyping today. A person spot-checks the output until the error rate earns their trust.
3. Checking and comparing
An invoice against its purchase order. A contract against your standard terms. An application against a document checklist. AI finds the matches and flags the mismatches, and a person decides what to do about each mismatch. This is the first rung where AI exercises real judgment, but a person still owns every decision that matters.
4. Drafting
Meeting minutes, monthly reports, routine replies, board-pack summaries. AI produces the first draft, and a person approves it before anything goes out. The time saved is large, and nothing leaves the organisation unreviewed.
5. Following up
Reminders on commitments from meetings. Chasing overdue actions. Flagging contract renewals before the deadline. Here AI starts acting on its own, but only within rules agreed in advance: whom it may remind, how often, and when silence is escalated to a person.
6. Running repeatable processes
Validate a monthly data file, compute the figures, produce the report. Process a batch of documents end to end. AI interprets the inputs, while the steps that must be exact, such as calculations and matching, are run by code. Every run is recorded: who started it, what went in, what came out.
7. Working with partners
The hardest work often crosses organisations. A consultant's proprietary scoring algorithm needs to run on your data. An assessor needs to check records you cannot share. At the top rung, one party's proprietary algorithms run on another party's data, and neither side sees the other's part. Every run is still on record.
What moves you up the ladder
The obvious answer is a smarter model. It is the wrong one.
What moves an organisation up the ladder is evidence. When finance can see that the last two hundred invoice matches were correct, and that every exception went to the controller, it becomes reasonable to let small, exact matches go through on their own. When leadership can see that reminders went to the right people at the right time, it becomes reasonable to let AI chase without asking first.
Each rung needs three things before you climb it:
- Clear limits. What the AI may do on its own, what needs a person, and what is prohibited.
- A safe way to fail. Mistakes on this rung are caught before they cause harm, by a citation, a review step or an approval.
- A record. Enough evidence to judge whether the AI has earned the next rung.
Skip any of the three, and trust never builds. The organisation stays on the bottom rung, or worse, someone climbs too fast without a record and the first visible mistake ends the project.
How to start
- Pick one process that crosses systems and that people find tedious: invoice checks, commitment tracking, application reviews.
- Start on rungs one to three. Let AI find, extract and check, while people keep every decision.
- Measure from day one. How often was AI right? Which exceptions did people override, and why?
- Climb one rung at a time, and only when the record says it is safe.
Where Forge fits
We built Forge so that organisations can climb this ladder safely. Answers cite their sources. Personal details such as email addresses, phone numbers and ID numbers are masked before a question reaches any model. Programs execute the exact steps. Administrators set a privacy floor that no individual can lower. Every access, AI call and program run is recorded, so the evidence to climb the next rung builds up as the work is done. Forge covers the lower rungs and repeatable processes today, and we are building towards following up and working with partners.
The goal is not to remove people from the work. It is to give them back the hours they spend carrying it, while they keep the hard calls.
Which rung is your team on today, and what would it take to climb the next one?
