Answer
What should be automated before building an AI agent?
The record-keeping, the notifications and the predictable branches. An agent built on top of an unrecorded process has nothing to work from.
The capture of what happens, the routine notifications, and every branch you can already describe. An agent placed on a process with no records and no rules spends its run rediscovering things a database would have told it.
The sequence matters because an agent's cost and unreliability both scale with how much it has to work out. A process where the state is recorded, the categories are defined and the ordinary paths are automated leaves the agent with the part that genuinely requires judgement. A process where none of that exists leaves it establishing basic facts by inspection on every run, expensively and sometimes wrongly.
Capture comes first. What arrived, when, from whom, what category, what happened to it, when it was resolved. This is unglamorous and it is the foundation for everything else: it makes the process measurable, it makes automation testable against history, and it gives any later system a source of truth rather than a reconstruction. Businesses that build this first find several of their intended automations become unnecessary, because the visibility was the actual need.
Notifications come second and are cheap. The reminder that something is outstanding, the alert that a threshold was crossed, the daily summary of what is waiting. These have no model in them, no permission surface, and no failure mode worse than a missed message, and they resolve a surprising share of the problems that motivated the agent conversation.
The predictable branches come third. Anything you can already describe as a rule should be a rule before an agent is involved, both because rules are cheaper and because writing them forces the process to be specified. Frequently the exercise reveals that what looked like judgement is a decision nobody had written down, and once written the agent has nothing left to decide.
What should remain for an agent after all this is the part that resisted every previous step: the case that requires reading unstructured language and working out what it means, the investigation whose next step depends on what the last one found, the item that does not fit any branch. That is a smaller and better-defined job, and an agent on it is both more likely to work and much easier to evaluate.
There is a sequencing benefit beyond the technical one. Each of these steps produces value on its own and can be stopped after. A business that captures, notifies and applies rules has improved its operations whether or not it ever builds an agent, which is a considerably better position than one whose entire return depends on the last and least certain step.
An agent is expensive at doing what a lookup does instantly, and most processes have several of those in the way first.
Siddharth Sharma, Context Theory
Related questions
Is this not just delaying the interesting part?
It is delaying it and reducing what it has to do, which is the point. An agent built on an unrecorded process is being asked to do the capture, the rules and the judgement at once, and it will do the first two badly and expensively. The interesting part gets easier and cheaper in proportion to how much of the dull part is already done.
How do you know when the groundwork is enough?
When you can describe what remains as a single question requiring judgement. If the remaining work still contains several steps you could specify, those should be specified first. The signal is being able to state the agent's job in one sentence that does not include the word and.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Odds of qualifying a lead — replying within the first hour vs after it | 7× | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | An agent's cost and unreliability scale with how much it must establish by inspection, so a process with recorded state and defined categories leaves it only the judgement, while an unrecorded one has it rediscovering basic facts each run. | Counting how many tool calls in an agent run are spent establishing facts a record would have supplied. |
| Response | Building capture first frequently makes several intended automations unnecessary, because the underlying need was visibility rather than automation of the step. | Comparing the automation candidate list before and after arrival, category and resolution data become available. |
| Software | Writing the predictable branches as rules forces the process to be specified, and the exercise commonly reveals that apparent judgement was an unwritten decision, after which the agent has nothing left to decide. | Attempting to write the rule for a branch currently described as requiring judgement. |
| Buying behaviour | Each preparatory step delivers value independently and can be stopped after, so the business improves regardless of whether the agent is built, which is a different risk position from one where the whole return depends on the final step. | Assessing whether the capture and notification layers would be worth keeping if the agent were never built. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one