ROI Assessment
Most AI returns are claimed rather than measured. This works out what actually changed.
What is ROI assessment?
An ROI assessment measures what a deployed AI system genuinely returned, comparing outcomes after deployment against a recorded baseline from before it, and setting the result against the full running cost rather than the build cost alone.
The hard part is almost always the starting point. If nobody recorded how long the job took or how many were handled per week beforehand, any improvement figure is a reconstruction rather than a measurement.
Where that baseline is missing we say so plainly instead of working backwards to a number that flatters whoever commissioned the system.
We measure what moved: time per case, volume handled, error rate, staff hours redeployed, revenue attributable to faster response. Then we set it against full running cost to produce a net figure rather than a headline one.
Why clients bring us supplier claims
Duara assesses AI returns from Nairobi, usually for boards deciding whether to expand a programme and for managers asked to justify last year spending.
We are frequently asked to verify a supplier claim. Where the system under assessment is one we built, that conflict is stated on the first page, and we will recommend an assessor with no delivery relationship if genuine independence matters more than familiarity.
Who this is for
- You have AI systems already in production
- You are deciding whether to expand or stop
- You must justify last year spending
- A supplier is claiming a strong return
Industries we serve
- Contact centres
- Financial services
- Retail chains
- Logistics operations
- Healthcare administration
- Manufacturing
Problems this solves
ROI work is nearly always commissioned at a decision point rather than out of curiosity.
- Returns claimed by whoever built the system. The figures come from the team with the strongest interest in them looking good, using measures chosen after the fact.
- No baseline recorded. Nothing was measured before deployment, so any improvement claim rests on memory rather than evidence.
- Gross benefit presented as net. The saving is quoted without subtracting licence fees, usage charges and the staff time to maintain the system.
- An expansion decision with no evidence. The board is asked to fund phase two without anybody establishing whether phase one worked.
What we deliver
A measured comparison against a stated baseline, with confidence levels declared, net of full running cost, ending in a recommendation to expand, adjust or discontinue.
Measured change against a stated baseline, with gaps declared
Net return after full running costs, not gross benefit
Which parts of the deployment earned their cost and which did not
A recommendation to expand, adjust or discontinue
Confidence levels on every figure presented
A baseline method for whatever you build next
Features
- Baseline reconstruction where possible. Where no baseline exists we rebuild an approximate one from system logs or payroll records and label its confidence.
- Net rather than gross. Running costs are subtracted, including the usage charges that grow with the very adoption being celebrated.
- Component-level attribution. We identify which parts of a deployment returned value, because programmes are rarely uniformly good or bad.
- Confidence stated on every number. Measured figures and estimated figures are visually distinguished so nobody confuses one for the other.
- Baseline design for next time. You leave with a method for recording the starting point of the next project, which is the cheapest thing you can do.
Benefits
- Expansion decisions rest on evidence. Phase two gets funded because phase one worked, not because phase one is now sunk cost.
- Failing components get cut early. Knowing which parts did not earn their keep saves the running cost of carrying them indefinitely.
- Supplier claims become checkable. A verified figure is worth considerably more than an asserted one when negotiating a renewal.
- The next project starts measurable. A baseline method in hand means the following assessment is faster, cheaper and firmer.
Technologies we use
- System log analysis. Extracting throughput, handling time and error data directly from the systems rather than from reports about them.
- Payroll and rota reconstruction. Establishing staff hours before and after where no formal baseline was ever recorded.
- Cost reconciliation. Pulling actual provider invoices rather than quoted rates, because the two frequently differ.
- Attribution analysis. Separating change caused by the system from change caused by seasonality or other concurrent projects.
- Confidence scoring. A consistent method for labelling how firm each figure is, so weak numbers cannot masquerade as strong ones.
Measurement work rather than build work. We read from your systems and records and change nothing while doing it.
How the process works, step by step
An ROI assessment runs three to four weeks and ends with a recommendation your board can act on.
-
Step 1
Free scoping call
We establish what was deployed, when, and whether anything was recorded before it went live.
-
Step 2
Baseline recovery
We reconstruct the starting point from logs, rotas and payroll, and declare how confident that reconstruction is.
-
Step 3
Outcome measurement
We measure what actually changed across throughput, handling time, error rate and staff hours.
-
Step 4
Cost reconciliation
We pull the real invoices, including usage charges, to establish what the system genuinely costs to run.
-
Step 5
Findings and recommendation
You receive the net figure, the component breakdown and a recommendation to expand, adjust or stop.
Quality assurance and how we work
What we hold ourselves to when measuring returns.
- A missing baseline is reported as missing. We will not work backwards to a flattering figure. Where the starting point is unknown, the report says so.
- Net figures only. Every return is presented after full running cost. Gross benefit on its own is not a result.
- We declare it when we built the system. That conflict appears on the first page, together with a recommendation to use an independent assessor if it matters.
- Concurrent causes are separated. Change from seasonality or other projects is identified rather than quietly attributed to the system under review.
- We report failure as readily as success. A finding that the deployment did not pay is delivered plainly, because that is the finding worth having.
Frequently asked questions about ROI assessment
What does an ROI assessment measure?
What a deployed system actually returned, compared against a recorded baseline and net of the full running cost rather than the build cost alone.
We never recorded a baseline. Is measurement impossible?
Not impossible, but weaker. We reconstruct one from system logs, rotas or payroll and state clearly how confident that reconstruction is.
What if the finding is that it did not pay?
Then we report that. Knowing which parts failed to earn their cost is what makes the next decision cheaper.
Why insist on net rather than gross?
Because usage charges grow with adoption. A gross saving that ignores them can describe a system that loses money.
How do you separate the system effect from other causes?
Attribution analysis, comparing against seasonality and any other projects running concurrently, so change is not misassigned.
Will you assess a system you built yourselves?
We can, and the conflict is stated on the first page. Where independence genuinely matters we will recommend an assessor with no delivery relationship.
How long does the assessment take?
Three to four weeks, most of it spent recovering the baseline and reconciling real costs.
What counts as a benefit?
Time per case, volume handled, error rate, staff hours redeployed and revenue attributable to faster response. Each measured separately.
Do redeployed staff hours count as a saving?
Only if the time went somewhere useful. Hours freed that disappear into meetings are not a return, and we check.
Can you assess part of a programme?
Yes, and component-level attribution is usually the most valuable part, because programmes are rarely uniformly good or bad.
What is a payback period?
How long a system takes to return its build cost, calculated after running costs are deducted.
Should we assess before deciding on phase two?
Yes. Funding an expansion without establishing whether phase one worked is how programmes accumulate cost without benefit.
Do you use provider invoices or quoted rates?
Actual invoices. Quoted rates and real bills differ often enough that using the quote would undermine the whole exercise.
How do you handle figures you cannot verify?
They are labelled as estimates with a confidence level, and never presented alongside measured numbers as though equivalent.
Can this support a renewal negotiation?
Yes. A verified figure is considerably stronger than an asserted one when a supplier proposes renewal terms.
What do we get for future projects?
A baseline method, so the next deployment starts measurable and the next assessment is faster and firmer.
Who should take part from our side?
Finance, the operational owner, and whoever holds the supplier relationship. All three hold part of the picture.
Is this the same as a productivity audit?
They overlap. ROI assesses one deployment against its cost; a productivity audit looks at whether tools you pay for are used at all.
How often should returns be reassessed?
Annually for systems still running, and immediately before any decision to expand or renew.
What does the assessment cost?
It depends on how many systems and how recoverable the baseline is. The figure is agreed before work starts.
Glossary of terms
Measurement terms as they apply here.
- Baseline
- A recorded measure of performance before a system was deployed, used as the comparison point afterwards.
- Net return
- Benefit remaining after all running costs are subtracted, as distinct from gross benefit.
- Attribution
- Establishing how much of an observed change was actually caused by the system rather than by something else.
- Payback period
- How long the system takes to return its build cost, calculated net of running cost.
- Confidence level
- How firm a figure is, distinguishing measured numbers from reconstructed estimates.
- Redeployed hours
- Staff time freed by automation and what that time was actually spent on afterwards.
Related services in AI Consulting
Talk to us about ROI assessment
If a supplier has given you a return figure, send it with the assessment request. Verifying a specific claim is usually quicker than measuring from scratch.
Start your project