AI assessments: find what is worth automating
An AI assessment should reveal what to automate, what the work costs and how to test the payoff. See the deliverables, a savings example and a free starting point.
An AI assessment for a business is valuable when it turns a recurring problem into a decision: what to automate, what the work costs today, and what evidence would justify the investment. You should leave with a specific workflow, a credible calculation and a practical way to test it. The value is being able to decide what to build and why.
Think about the report that interrupts every Friday. Someone exports figures, checks dates, copies them into a template and asks a colleague why two totals disagree. The business already has software. It may even have several AI subscriptions. Yet the report still needs a person to push it through each step.
An assessment makes that work visible. It separates the part a system could handle from the decisions that need your team. That is where an interesting AI idea becomes a useful business project.
What should an AI assessment actually include?
The word assessment covers very different services. An AI readiness assessment might examine data access, staff skills and ownership across the company. A workflow assessment might focus on a single process. A model evaluation asks whether a particular model can perform a defined task reliably.
Know which one you are getting. A company-wide maturity score does not, by itself, tell the person preparing Friday's report what will change next week.
For a practical automation assessment, I would expect these deliverables:
| What you receive | What it lets you decide |
|---|---|
| A map of one real workflow | Where the work starts, passes between people or tools, and finishes |
| A current-work baseline | How much hands-on effort the process takes in a defined period |
| A proposed automation | Which steps a system could take over and what stays with a person |
| A priority and its reasoning | Why this task deserves attention before the other ideas |
| An assumptions and access list | What is known, what is estimated and what needs checking |
| A test and success criteria | What evidence would make the first version worth keeping |
This emphasis on context has a sound basis. The NIST AI Risk Management Framework calls for defining business value, expected benefits and costs, application scope and human oversight. It also separates mapping a system from measuring its performance. A useful commercial assessment should make those distinctions understandable without handing you a policy manual.
Why assessments reveal opportunities a software demo misses
A demo begins with something the product can do. Your assessment begins with work that needs doing.
Take an online retailer answering product compatibility questions. The visible task is writing replies. The harder part might be finding the right specification and checking that it belongs to the exact product variant. A faster writing model will not repair an outdated source of truth.
The useful automation could retrieve the approved specification, draft an answer with a source reference, and send uncertain matches to a person. That is a buildable workflow. It also explains why simply buying another chat subscription may leave the original problem intact.
A manufacturer preparing quality reports has a different constraint. If departments calculate the same metric differently, the first version needs agreed definitions. The automation can then collect the inputs, apply those definitions and flag missing records. Its job is more precise than “use AI for reporting.”
An agency chasing approvals might need a system that gathers feedback, identifies the current version and reminds the named approver. Creative approval still belongs to a person. The assessment finds the preparation and coordination that can be removed around that decision.
These are illustrative opportunities, not promised integrations. In each case, the important questions concern the task, the available information and what a usable output must contain.
How to calculate the value without inventing savings
Start with a period you can describe. Count either the total staff time spent on the task or the number of items multiplied by average hands-on time per item. Include everyone involved. Avoid counting the same effort twice.
Waiting time belongs in a separate measure. A purchase request sitting untouched for two days has a delay problem, but that does not mean someone spent two working days processing it.
Then distinguish three figures:
- Current workload: time the task takes now, supported by records, a timed sample or a labelled estimate.
- Potential time recovered: current workload minus all the staff effort the proposed process would still require.
- Financial effect: what that recovered capacity is worth, less setup and operating costs, with any actual cash or revenue effect identified separately.
A worked example with visible assumptions
In the illustrative stock and sales workflow map, the planning inputs describe 48 reports over four weeks, taking an estimated 25–35 staff minutes each. That produces a current workload of 20–28 staff hours. The example then supplies a target of 6–8 total staff hours for the same period, including review and upkeep.
The conservative end subtracts the higher remaining effort from the lower current workload: 20 minus 8 equals 12 hours. The other end subtracts the lower remaining effort from the higher workload: 28 minus 6 equals 22 hours. The resulting scenario is 12–22 staff hours recovered over four weeks, if the target is achieved.
At the example's assumed staff cost of 600 NOK per hour, that is 7,200–13,200 NOK of equivalent capacity before setup and running costs. Source: illustrative workflow map, four-week planning scenario.
This is useful because the assumptions are visible. It is not evidence that a real business has saved that amount. A paid employee's salary does not disappear when a task becomes faster. The commercial benefit might be capacity for more orders, a reduced backlog or less overtime, but each of those needs its own evidence.
If you do not know the future workload, leave it unknown. A trial should establish it. An assessment that produces a confident saving from missing information is giving you a number to admire, not a decision you can trust.
Why the same automation can produce different results
Research illustrates why a universal saving is a weak starting point. In Generative AI at Work, researchers studied an assistant used by customer support agents and found that productivity effects differed substantially across workers. That setting does not supply a savings percentage for your warehouse, studio or finance team. It does support checking who uses the tool and where it changes their work.
For your process, collect normal examples and a few difficult ones. Ask the people doing the work what they check before considering it finished. Watch for effort that is easy to miss: finding an attachment, fixing a customer name, interpreting an exception or copying the approved output into another system.
Then compare complete results. A draft that takes seconds but needs extensive correction may be less useful than a slower output that needs a quick check. Measure quality alongside time, and record cases that failed rather than excluding them from the result.
What happens after a useful assessment?
You should be able to describe the first version in ordinary language. For the reporting example: collect the agreed exports, check that the periods match, prepare the recurring report and flag missing or inconsistent figures for review.
Before building, establish whether the required tools can actually exchange that information, who controls access and who owns the finished report. The assessment should expose those questions early enough to shape the scope and quote.
A good next step is a bounded pilot. Run a representative batch through the proposed process, measure the complete effort, compare the output with your quality criteria and decide whether to expand it. Sometimes the conclusion is to adjust an existing software feature. Sometimes it is a small custom automation. Sometimes the task needs clearer inputs before a build will help.
That is also the distinction between a free self-check and deeper analysis. My free bottleneck check uses your answers to create a workflow map, a proposed automation where appropriate and a prioritised next step. It can calculate workload from the figures you supply. It cannot inspect your private systems, verify an integration or prove that a target will be achieved.
A deeper assessment brings real examples, tool access and the people involved into the investigation. Those extra inputs should resolve questions that materially affect what gets built.
What does leaving the process unchanged mean?
If the same volume of work returns and the process stays the same, the effort is likely to return with it. The practical cost is the useful work your team continues to postpone while handling that recurring task.
Name that alternative. Would recovered time go towards clearing overdue orders, improving customer replies, checking quality or preparing proposals sooner? This gives the assessment a purpose beyond making a large number appear on a screen.
It also gives you permission to deprioritise a weak opportunity. An infrequent task with little burden may deserve less attention than a daily handoff that repeatedly delays the next person. Compare the cost of changing the process with the cost of leaving it alone.
Start with the task you would most like to stop repeating
You do not need an inventory of every system in the company. Start with one recognisable task, a recent period and an honest account of what happens from beginning to end. A description in your own words is often enough to reveal which details need investigating next.
Run the free bottleneck check to get a first map without an email gate. You can leave unknown figures blank and save the result. For examples of small, concrete builds, read four workflow automations for a small team.
If the result points to something worth pursuing, bring the map to a free call. We can turn the most useful opportunity into a first version, identify what needs checking and agree how to judge the result before you commit to implementation.
Questions people ask
What is an AI assessment for a business?
An AI assessment examines a business process, its inputs, tools, people and current workload to identify useful applications of AI or automation. A useful result names a candidate workflow, shows the evidence behind its value and explains how to test it before committing to a build.
What is the difference between an AI readiness assessment and an automation assessment?
A readiness assessment examines whether a business has the data, access, skills and ownership to use AI effectively. An automation assessment looks more closely at a particular workflow and its potential value. A practical project may need both, but the scope should be explicit.
Can a free AI assessment tell me exactly how much I will save?
A self-assessment can calculate current workload from the figures you enter and compare it with a target you supply. It cannot verify future savings or software compatibility. A representative trial, including review, corrections and upkeep, is needed to establish the actual result.
Do I need to know my numbers before starting?
No. You can start with a task you recognise and leave unknown figures blank. The free bottleneck check can still suggest a workflow and next step. Estimates should be labelled, and the next useful action may be timing a few representative examples.
Is an AI assessment only useful for service businesses?
No. Retailers, manufacturers, software companies, agencies and internal departments all have workflows worth examining. The assessment should focus on a task and the people doing it, rather than assume every business sells services or handles enquiries in the same way.
Sources
- 1.NIST: AI Risk Management Framework core · Guidance on business context, intended benefits, costs, scope, oversight and measurement; not an endorsement of this assessment
- 2.Brynjolfsson, Li and Raymond: Generative AI at Work · Research on an AI assistant in customer support; effects varied across workers and do not predict results for another business
- 3.Ivar Knutsen: illustrative stock and sales workflow map · Source for the worked example; all figures are planning assumptions, not measured client savings
Written by Ivar André Knutsen
I build and run AI systems, internal tools and workflow automation. You work directly with me from the first conversation through implementation and support. About Ivar
Want this looked at in your business?
We look at where you want the business to go, what is slowing you down and where AI could make a useful difference. You get a clear recommendation: a tool to try, a focused automation, a broader system or a closer look at the process. Any build is scoped and quoted before work starts. Free, no obligation.
Single automations are quoted on the call: a setup fee plus a monthly retainer to run them. Full systems start at $4,500, fixed scope, fixed price.
Book a free 30-minute call