AI Cost Analysis
The build price is rarely the number that matters. What counts is the monthly cost afterwards.
What is AI cost analysis?
AI cost analysis establishes the total cost of owning an AI system: the build, the licences, the per-use charges that grow with adoption, the hosting, the monitoring and the staff time to maintain it, compared against what the task costs you today by hand.
AI carries a cost shape most businesses have not budgeted for before. A great deal of it is charged per use, so the more customers you serve the larger the bill becomes. A quiet month is cheap and a busy month is not.
Add licences, hosting, monitoring and the staff hours to keep it working, and the annual figure frequently bears no resemblance to the original quote.
The comparison that decides everything is against your current manual cost. Without that number you cannot tell whether automation is genuinely cheaper or merely newer, and most proposals never establish it.
Who usually hires us for this
Duara models AI costs from Nairobi for businesses comparing supplier quotes, and for finance teams asked to approve spending they cannot yet interrogate.
We are frequently hired to examine somebody else proposal. When our own work is among the options being compared, that appears at the top of the report, because a cost analysis from an interested party is worth reading differently.
Who this is for
- You are comparing quotes that are not like for like
- Your finance team must approve the spend
- A pilot cost jumped once real volume arrived
- Nobody has costed the current manual process
Industries we serve
- Retail and eCommerce
- Contact centres
- Financial services
- Logistics
- Healthcare administration
- Professional services
Problems this solves
Cost work usually starts when a bill arrives larger than expected, or when finance refuses to sign something they cannot break down.
- Quotes that are not comparable. Two suppliers price different scopes with different inclusions, so the cheaper number is not actually the cheaper option.
- Costs that grow with success. Per-use pricing means the system gets more expensive precisely as it starts working, which nobody budgeted for.
- Running costs omitted entirely. The proposal covers the build and stops, leaving licences, hosting, monitoring and maintenance unquantified.
- No manual baseline. Nobody costed doing the work by hand, so there is nothing to compare the automation against.
What we deliver
A total cost of ownership model covering build, run and maintenance, projected at realistic volume, set against the measured cost of your current manual process.
Total cost of ownership across build, run and maintenance
How the cost changes as usage grows, not just at pilot size
What the task costs you today, done manually
Where the pricing model exposes you and how to cap it
A like-for-like comparison of competing quotes
A breakdown your finance team can challenge line by line
Features
- Volume modelling. Costs are projected at genuine transaction volumes rather than the pilot numbers that flatter every proposal.
- Manual baseline measurement. We cost the current process in staff hours, so the comparison has two real sides rather than one.
- Quote normalisation. Competing proposals are restated onto identical scope and inclusions before any figure is compared.
- Cap mechanisms identified. Usage limits, caching, cheaper model routing and self-hosting thresholds, with the point at which each becomes worthwhile.
- Sensitivity analysis. What happens to the annual figure if volume doubles, or if provider pricing rises, both of which happen.
Benefits
- Finance can approve with confidence. A breakdown that survives line-by-line challenge moves faster than a headline figure that does not.
- No surprise at scale. The month the system succeeds is not the month the bill becomes a problem.
- Quotes become genuinely comparable. Restating scope reveals which supplier is actually cheaper, which is often not the one with the lower number.
- You know when to walk away. Sometimes the analysis shows the manual process is cheaper, and that is a valuable answer.
Technologies we use
- Usage projection. Modelling transaction volumes from your own history rather than from supplier assumptions.
- Provider pricing analysis. How each provider charges, including the tiers and thresholds that change the effective rate.
- Caching and routing economics. When reusing previous results or sending simple work to cheaper models actually pays.
- Self-hosting threshold analysis. The volume at which running your own infrastructure becomes cheaper than per-use pricing.
- Staff time measurement. Costing the manual process in hours so the comparison rests on measurement rather than estimate.
This is financial modelling applied to technical systems. The only tooling involved is what is needed to measure your current volumes accurately.
How the process works, step by step
Cost analysis takes two to three weeks, most of which is establishing your real volumes and current manual cost.
-
Step 1
Free costing call
We establish what is being proposed, who has quoted, and whether anyone has costed the current manual process.
-
Step 2
Volume measurement
We establish your genuine transaction volumes from your own records rather than from projections.
-
Step 3
Manual baseline costing
We cost the existing process in staff hours so there is a real number to compare against.
-
Step 4
Model and normalise
We build the ownership model and restate any competing quotes onto identical scope.
-
Step 5
Finance walkthrough
We present the breakdown and take challenges directly from whoever has to approve it.
Quality assurance and how we work
The commitments attached to cost work.
- We declare our own interest first. If Duara work is among the options being compared, that appears at the top of the report, not in a footnote.
- Every figure has a source. Costs trace to a provider price list, a measured volume or a staff hour rate, never to a rule of thumb.
- Volumes come from your records. Projections are built from what you actually processed, not from what the proposal assumes you will.
- We model the bad case too. What the annual figure becomes if volume doubles or provider pricing rises is stated rather than omitted.
- We will tell you not to buy. Where the manual process is cheaper at your volume, that is the conclusion, and we have written it before.
Frequently asked questions about AI cost analysis
What does AI cost analysis cover?
The full cost of owning a system: build, licences, per-use charges, hosting, monitoring and maintenance, compared against your current manual cost.
Why do AI costs rise when the system works well?
Most AI is charged per use. Serving more customers means more usage, so cost grows with success in a way fixed-licence software does not.
Can the cost be capped?
Usually. Usage limits, caching repeated answers, routing simple work to cheaper models, or self-hosting once volume justifies it. Best designed in early.
What is a token and why does it matter?
The unit most language model providers bill by, roughly a fragment of a word. It matters because it is what your invoice is calculated from.
How do you compare quotes from different suppliers?
By restating them onto identical scope and inclusions first. The lower headline figure is frequently not the cheaper option.
Why measure the manual process?
Because without it you cannot say whether automation is cheaper. Most proposals never establish this number.
What is self-hosting and when is it worth it?
Running the models on infrastructure you control. It becomes cheaper past a certain volume, and we calculate where that point sits for you.
Do you model what happens if volume doubles?
Yes. Sensitivity analysis is part of the work because volume growth is the most common cause of budget failure.
What if provider pricing changes?
We model that too. Pricing has moved repeatedly, usually downward, but assuming it will is not a plan.
How long does the analysis take?
Two to three weeks, mostly spent establishing real volumes and costing the current manual process.
Will you assess a proposal from another supplier?
Yes, frequently. If our own work is among the options, that is declared at the top of the report.
What is caching in this context?
Reusing a previously generated answer rather than paying to produce the same result again. It can substantially reduce recurring cost.
What is model routing?
Sending straightforward requests to a cheaper model and reserving the expensive one for genuinely difficult work.
Do hidden costs usually appear?
Monitoring and staff maintenance time are the two most commonly omitted, and together they are rarely trivial.
Can you analyse a system already running?
Yes, and the numbers are better because actual consumption is visible rather than projected.
What if the analysis says do not automate?
Then that is the finding. At low volumes the manual process is often genuinely cheaper, and we have written that conclusion before.
Who needs to take part in a cost review?
Whoever holds the budget and whoever knows the real volumes, usually finance and operations together.
Does this include implementation cost from us?
Only if we are being considered for the work, and then it is presented alongside the alternatives with our interest declared.
How accurate are the projections?
They carry stated confidence. Volumes measured from your records are firm; provider pricing changes are explicitly uncertain.
What does the analysis itself cost?
It depends on how many options and systems are in scope. The figure is fixed before we start.
Glossary of terms
Cost vocabulary as it applies to AI systems specifically.
- Total cost of ownership
- Build cost plus everything required to keep a system running, including licences, usage and maintenance.
- Per-use pricing
- Charging by unit of consumption, so the bill rises directly with adoption rather than staying fixed.
- Token
- The unit most language model providers bill by, roughly corresponding to a fragment of a word.
- Caching
- Reusing a previous result instead of paying to generate the same answer again.
- Model routing
- Sending simpler requests to a cheaper model and only difficult ones to an expensive one.
- Baseline cost
- What the task costs you today under the current manual process, measured in staff hours.
Related services in AI Consulting
Talk to us about AI cost analysis
Bring the quotes you have already received. Restating them onto the same scope is usually the fastest useful thing we can do.
Start your project