Where should a mid-sized company start with AI so the budget is not wasted?

Start with the work, not the tools. Before a plan and a budget are signed off, find out how the work actually gets done from the people who do it, put a cost on each problem they describe, and rank the list with the largest return at the top. Only then decide what to automate first and what to leave alone. Most AI budgets skip this step. They pay to automate the documented process, and the documented process is not the one running.

Why the budget lands on the wrong problem

The failure rate is covered in Why AI transformation fails: nobody knows how the work actually happens. The question for an owner about to spend is different: why does the money land on the wrong problem? Three patterns show up in the research.

Spending before understanding. IBM's 2025 survey of 2,000 CEOs found that 64 percent invest in some technologies before they understand the value, because they fear falling behind. Only 25 percent of AI initiatives had delivered the expected return.

Spending where the result is easy to see rather than where it is large. MIT NANDA's 2025 study found that about half of generative AI budgets go to sales and marketing, while the better returns came from back-office automation. The authors put this down to easier attribution, not actual value. A sales dashboard makes a good board slide. The credit-hold queue does not, and that is where the hours go.

Misreading the problem. RAND's 2024 study of failed AI projects found the most common root cause was that stakeholders misunderstood or miscommunicated what problem needed solving, so models were built for the wrong metric or did not fit the workflow around them. A close second: chasing the latest technology instead of a real problem.

Together they produce shallow AI: AI bolted onto work nobody described. It passes the demo, because the demo follows the manual. It fails in month three, because the team never followed the manual either.

The manual is not the map

Every mid-sized company runs on tribal knowledge. The official order-to-cash process has five steps. The real one has a pricing check that lives in one person's head, a spreadsheet that overrides the ERP's lead times, and a customer waved through credit control because someone once decided they could be trusted. None of that is written down. All of it is where the time goes.

These workarounds are the fixes people made when the system did not fit the job. They are invisible to anyone planning a budget from the process documentation and the org chart. Only the people doing the work can describe how work actually gets done, and they will only do so privately.

So the map is built from the bottom: everyone who touches the work, asked separately, with the contradictions between their accounts kept. Where two teams describe the same handoff differently, that is usually the most expensive step in the company.

Rank by what each problem costs

A map does not tell you where to spend. The ranking does. For each problem people describe, three numbers are enough to start: hours (how many people spend how long on it each week), errors (how often it goes wrong and what a mistake costs), and waiting (how long work sits in a queue and what the delay costs in cash or customers).

Take Acme, a 220-person building-products distributor. Its leadership expected the first AI budget to go into a sales assistant, because that was the demo they had seen. The people doing the work said something else.

The sales assistant sat at the bottom. The re-keying, the quote bottleneck and the proof-of-delivery chase together took more than a hundred hours a week and held up cash, and none of them appeared in the process manual, because the manual did not know they existed. The list changes the budget, not the other way round.

What to automate first, and what to leave alone

With a ranked list, the first decision is straightforward. Start where the return is largest and the routine case has clear rules: high volume, repeated daily, most instances handled the same way. At Acme, the emailed orders.

Just as important is what to leave alone, at least at first:

• Work that is mostly judgement. Stack Overflow's 2025 survey found that 66 percent of developers name AI output that is "almost right, but not quite" as their biggest frustration. Almost right is fine when a rule catches the miss. It is expensive on a credit decision.
• Processes about to change anyway. Automating a step a new system will remove is money spent twice.
• Problems that are really a people or policy problem. Acme's quote bottleneck was one estimator holding pricing rules nobody had asked him to write down. The fix started with capturing those rules, so the company no longer depended on one person, and only later with software.
• Low-volume, bespoke work. There is little routine in it to remove.

PwC's 2026 predictions put the balance plainly: technology delivers about 20 percent of an initiative's value, and the other 80 percent comes from redesigning the work. You cannot redesign what you have not mapped, and you should not redesign what you have not costed.

What the first budget should buy

The map and the ranked list, not the tool. Before a build budget is approved, an owner should hold three things: how work actually gets done in each function, with the workarounds and single points of failure named; each problem costed in the company's own hours and money, ranked; and a decision on the first one to three automations, with a written list of what is being left alone and why.

Then build in that order and measure the saving on the books. The build sequence is in How should a mid-sized company start becoming AI-native?. This guide is about the step most companies skip before it.

Our view

Our view, and it is ours: the AI budget should not be planned from the org chart, the process manual or a vendor demo. It should be planned from a world model of the company, built from the people who do the work.

Transcript was built for this. It holds private conversations with everyone in the company at once, keeps the contradictions between accounts, and turns them into a world model: how each process actually runs, what each problem costs, and who everything depends on. The model gets richer as the company keeps talking to it, and over time it starts to show where the next problem will appear, not only where the last one was.

What happens next is the company's choice. Our engineers can build the first automations with the team, in the order the ranking gives. Or the company can take the model and get on with it. Either way, the budget is spent on the work that is actually there.

Questions people ask next

How do I avoid wasting money on AI?

Spend the first money on finding out how the work actually gets done and what each problem costs. Approve a build budget only against a ranked list. If a proposal cannot say which hours it removes, it is not ready to fund.

What should a small or mid-sized business automate first?

The problem at the top of the cost ranking that also has clear rules for the routine case. It is usually a high-volume back-office task such as order intake, invoice matching or document chasing, not the demo that prompted the budget.

Why should I map the process before automating it?

Because the documented process is rarely the one running. Automating the manual gives you a faster version of a process nobody follows, and the team routes around it within weeks. Mapping from the people doing the work shows the real steps, the workarounds and the single points of failure.

Do we need an AI consultant, or can we do this ourselves?

You can do the mapping yourselves if someone has the time to talk to everyone privately and the standing to hear the awkward parts. In a company of 50 to 500 people that is hundreds of conversations, and staff describe workarounds honestly only when their manager cannot read the answers. Whoever does it, insist on the ranked list before anyone proposes a build.

What business processes should you not automate?

Work that is mostly judgement, processes about to change anyway, problems that are really one person holding knowledge nobody wrote down, and low-volume bespoke work. Each needs a different fix first.

References

1. IBM Institute for Business Value, IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles (May 2025). https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
2. MIT NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025), as reported by Fortune, MIT report: 95% of generative AI pilots at companies are failing (18 August 2025). https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/ Report PDF: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
3. RAND Corporation, Ryseff, De Bruhl and Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (August 2024). https://www.rand.org/pubs/research_reports/RRA2680-1.html
4. Stack Overflow, 2025 Developer Survey, AI section (2025). https://survey.stackoverflow.co/2025/ai
5. PwC, 2026 AI Business Predictions (2025). https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html