Every week someone asks whether AI can “handle” a process. The honest answer is usually not yes or no. It is: which parts, under what rules, and who owns the exceptions when the model is wrong.
As a Managed Intelligence Provider, we do not start with a model. We start with the work — the inputs, the decisions, the hand-offs, and what “done” looks like on a normal Tuesday.
Here is a filter you can use before anyone builds anything. It will save you from decorating a fragile workflow with clever language.
Start with the shape of the work
AI helps most when the same kind of information shows up often, and the same kind of judgement gets applied most of the time. That is not the same as “this job is hard.” Hard, rare, high-stakes judgement may stay human forever. That is fine.
Ask five questions:
- Is there volume? Dozens or hundreds of similar cases beat a handful of unique ones. Models earn their keep on repetition.
- Is there sameness? Can you name the rules that cover most cases? If every instance is a special, you do not have an AI problem — you have a process-design problem.
- Are the inputs digitable? Emails, PDFs, forms, tickets, structured fields. Work that only exists as hallway conversation is not ready.
- Is “done” clear? Extracted fields, a draft reply, a routed queue, a pass/fail with a reason. Vague goals produce vague outputs.
- Can a human catch the miss? Consequential steps need an exception path with ownership — not a silent wrong answer.
If three or more answers are soft, fix the information path first. Document the real steps. Structure the intake. Put the source of truth in one place. Then revisit AI.
Where AI usually fits
Think of AI as labour for reading, classifying, extracting, drafting, and triaging — the middle of the process where humans currently re-type, hunt, and assemble.
- Pulling fields from invoices, certificates, or intake packs into a system
- Sorting inbound work into queues with a reason code
- First drafts of replies, summaries, or packs that always start from the same ingredients
- Matching records that almost agree and flagging the ones that do not
- Checking completeness before a case is allowed to move
Notice what is missing from that list: unsupervised decisions that move money, change a customer relationship, or commit the firm with no review. Those can be assisted. They should not be abandoned to a model.
Where it does not belong (yet)
Skip AI — or keep it far from the consequential step — when:
- Only one person knows the exceptions, and nothing is written down
- Inputs are chaotic and nobody agrees what a complete case looks like
- The “process” is actually negotiation, relationship, or rare judgement
- There is no way to spot or reverse a wrong answer quickly
- Success would be measured by demos instead of cycle time, rework, or accuracy
In those cases the right move is process intelligence and information architecture: make the work visible and transferable. Applying a model to undocumented single-person work just accelerates the mess.
What good looks like in production
Good is not a chatbot on the intranet. Good is a thin slice of real work running under clear rules.
Shadow before you trust
Run the model alongside humans first. Compare outputs. Measure agreement and the kinds of misses. Promote only what earns it.
Exceptions are the product
Uncertain cases go to a queue with a reason, not a confident guess. Someone owns that queue. Layouts change. Suppliers invent new formats. Managed operations means the edge cases have a home.
Leave a trail
Every automated step should show what it saw, what it decided, and why. If you cannot audit it, you cannot improve it — and you cannot defend it when something goes sideways.
Keep humans on the consequential step
Drafts, extracts, and triage can be machine-led. Approvals, commitments, and irreversible actions stay human until the evidence says otherwise.
How Kultiv decides
We watch a live case end to end before we talk about models. We cost the hours in rework, waits, and re-typing. The case for AI is a process case — volume, sameness, clear done — not a technology case.
When a slice fits, we put the smallest useful piece into production inside your environment, often shadowing first. Then we run it: monitoring, exceptions, accuracy, and a written monthly review. That is managed operations, not a pilot that dies in a slide deck.
If you are staring at a process and wondering whether AI belongs there, start with the five questions above. The answer will usually tell you whether to map, systematize, assist — or leave well alone for now.
Bring us the process you are tempted to “AI”
A discovery call is about forty minutes and costs nothing. We will help you tell whether AI belongs in that work — including when the honest answer is not yet.