Free Tool
Enterprise AI assistant payback calculator
“Where is that information?” gets asked dozens of times a day in every company. This tool prices that question and estimates how fast an assistant pays for itself.
The calculation rests on one comparison: how many minutes it takes to find an answer today, and how many it would take with an assistant. The difference is multiplied by the number of questions the assistant can actually cover, then converted into annual hours and money. You supply the coverage share — there is no invented “AI productivity rate” on screen. The default setup cost sits inside KodDelta’s published RAG band of $5,000 - $12,000.
This is an estimate, not a quote. Scope and price are fixed after a discovery call. Let us look at your documents →
How the calculation works
- Hours spent today = questions per month × today’s average minutes ÷ 60 × 12. This is the total time spent working out where information lives.
- Questions covered = questions per month × the coverage share you entered. The remaining questions are still resolved by people, exactly as before.
- Hours with the assistant = (covered questions × assistant minutes + uncovered questions × today’s minutes) ÷ 60 × 12. Note that uncovered questions keep their original duration — the assistant does not speed them up.
- Hours recovered = today’s hours − hours with the assistant. Annual value = hours recovered × fully loaded hourly cost.
- Net annual value = annual value − (monthly running cost × 12). Payback = setup cost ÷ (net annual value ÷ 12), in months.
There is no other coefficient in the model. In particular, we do not use: a benchmark average lookup time, any “AI raises productivity by X%” rate, or hypothetical revenue growth from better quality. Each of those would produce a bigger number on screen, and none of them has been measured in your company.
Filling in “questions per month”
Most companies have never measured this, but it is easy to estimate. Add up three channels:
- Internal messaging. “Where is that customer’s contract”, “what was the tolerance on this part”, “can you find last year’s quote”. Review one week of chat, count them, multiply by four.
- Phone and in person. Questions from the warehouse to production, from the field to head office. This is usually the largest channel and it leaves no trace anywhere.
- Self-directed searching. Time spent digging through folders, email and old files alone. Because nobody is asked, it is invisible — and it is usually the longest line of all.
Do not inflate the figure. A small, defensible number produces a far more useful result than a padded one, because you can actually base a decision on it.
Coverage share: the most critical and most overestimated field
Only one thing determines how many questions an assistant can cover: whether the answer is written down somewhere. However good the model, it cannot produce information that exists nowhere — and if it does produce it, that is precisely the problem.
In practice the question flow splits three ways. First, questions whose answers sit in a procedure, contract or product file; these can move over directly. Second, questions whose answers exist as data in a system but not in any document (“what were this customer’s last three orders”); these need the assistant connected to the system and count as extra scope. Third, questions requiring judgement and decision; these stay with people, and they should.
Split your own flow that way and enter only the first bucket as a percentage. A modest result is not bad news — it tells you accurately where the assistant will earn its place.
Do not read the payback period optimistically
The number of months on screen is the best case, because three realities sit outside the calculation. First, document preparation: a scattered archive with version confusion and unclear ownership has to be assembled first, and that effort is mostly on your side. Second, the settling-in period: for the first few weeks the team both asks the assistant and verifies the old way, so nothing gets faster. Third, scope growth: as the assistant proves useful, new document sets get added and the setup cost turns from one-off into a phased investment.
There is also one item outside the calculation that runs in your favour: reaching the right answer quickly does not only save time, it reduces work done on wrong information. The cost of a part machined to an outdated tolerance, or a quote sent from a superseded price list, never shows up in this tool.
The technical frame of an assistant build
The assistants KodDelta builds follow three rules. Source attribution is mandatory: every answer states which document and which section it came from; an answer that cannot be verified is not given. Permission boundaries hold: the assistant will not answer from a document the asker is not allowed to see, and access rights are inherited from the source system. Unknown questions are refused: when no source is found the model does not guess, it says the documents do not cover it. Without those three an assistant is fast but untrustworthy — and an untrustworthy assistant stops being used within weeks.
To see the AI scope alongside everything else, tick the AI module in the budget and timeline estimator. If the assistant also needs to reach system data rather than just documents, the integration complexity score measures that part separately. For the subject in prose see our AI process automation page.
Frequently Asked Questions
What does a RAG assistant actually do?
RAG (retrieval-augmented generation) means the model answers from your documents rather than from its own memory. Procedures, quotes, contracts, product files and past records are indexed; when a question arrives the relevant passages are retrieved first and the answer is written only from those passages, with the source shown. That makes the answer auditable — no sentence is produced without a source behind it.
What should I enter as the coverage share?
That field is deliberately your own estimate and the default is meant to be changed. Questions whose answers are written down in a document can move to the assistant; questions requiring interpretation, negotiation or judgement stay with your team. Split your own question flow into those two buckets and derive a rough percentage. We offer no industry average, because this share depends entirely on how well your documents cover your work.
How is the setup cost derived?
KodDelta’s published RAG assistant band is $5,000 - $12,000, and the default in the field sits inside it. What moves the figure is the number and format of documents, the complexity of access permissions, and the depth of accuracy testing. The assistant can also arrive as part of a full platform delivery rather than as a standalone project.
Why does the running cost default to zero?
Because without knowing your model provider, your usage volume and your hosting preference we cannot know that figure, and we are not willing to invent it. The field is left open for you: enter a monthly estimate from your own provider’s pricing and the net benefit and payback period adjust accordingly.
What happens when the assistant does not know?
A correctly built RAG assistant does not answer a question it has no source for; it says the documents do not cover it and routes the question to the right person. Guaranteeing that behaviour — through hard boundaries and source attribution — is the most important part of the build. Auditability comes before speed.
Let us pick the first document set together
Thirty minutes covers scope, boundaries and how accuracy will be tested.