How should a mid-sized company start becoming AI-native?
Start by mapping how work actually happens, pick one high-volume process, redesign it around AI, build the fix, and measure the saving on the books.
A mid-sized company should start by finding out how its work actually happens, from the people doing it, and then redesign one high-volume process around AI before touching anything else. Buying tools first is the common mistake and the expensive one. The order is: map, pick, redesign, build, measure, and only then repeat.
Step 1: Map the work as it actually runs
Every company has two versions of its processes. The documented one, and the one running in the shared inbox and the reconciliation spreadsheet. Only the second one costs money.
Mapping means asking the people doing the work: where does it arrive, who touches it, where does it wait, what goes wrong. Management interviews miss most of this, because managers describe the designed process. Doing it by hand takes months, which is why most companies skip it and buy a tool instead. IBM's 2025 CEO Study (May 2025) shows where that leads: half of 2,000 CEOs said the pace of AI investment had left them with disconnected, piecemeal technology.
We built Transcript to do the mapping at scale: private conversations with everyone at once, contradictions flagged, and a world model of the company with every problem valued in its own hours and cost figures.
Step 2: Pick one process on volume and rule clarity
Start with the one that has high volume, clear rules for the routine case, and a person currently doing that routine case by hand.
| Candidate process | Volume | Rule clarity | Routine case done by hand today | Start here? |
|---|---|---|---|---|
| Supplier invoice handling | High | High | Yes | Yes |
| Order intake from email | High | Medium | Yes | Yes |
| Month-end reconciliation | Medium | High | Yes | Yes |
| Customer complaints | Medium | Low | Yes | Later |
| Hiring and screening | Low | Low | Partly | No |
The last row has a regulatory reason. Under the EU AI Act (Regulation (EU) 2024/1689, in force since August 2024), AI used in employment decisions is classed as high-risk and carries formal obligations, while the vast majority of AI systems, including back-office automation, sit in a minimal-risk category with none.
Step 3: Redesign the process before building anything
Take the mapped process and ask of each step: should this exist. An approval that takes four days usually exists because nobody removed it. Then define three things for what remains: the rule for the routine case, what counts as an exception, and who owns the exception with what time window.
Only now does the technology question arise. PwC's 2026 AI Business Predictions state that technology delivers about 20 percent of an initiative's value, and 80 percent comes from redesigning work so that agents handle routine tasks.
Step 4: Build the routine case, keep people on the exceptions
The failure we see most often is a handoff. One team maps and recommends. A different team, or nobody, builds.
Build the routine case first and route exceptions to the named owner. Run the old process in parallel for a period and compare. Stack Overflow's 2025 Developer Survey gives the reason: 66 percent named AI output that is "almost right, but not quite" as their biggest frustration. Almost right is fine for the routine case, where a rule catches the miss. It is expensive on the exception.
Step 5: Measure on the books, then repeat
Baseline the process cost before the build, in the CFO's figures. After the build, measure hours removed and cost removed for at least a quarter. If the figure does not survive contact with the ledger, the project is not finished.
Our view: skip the strategy phase
Most advice says to begin with an AI strategy, a Chief AI Officer, and a data foundation. We disagree for companies of this size. Eurostat's 2025 survey (December 2025) counts 30 percent of EU medium-sized enterprises as AI users, yet in our experience very few can name a process whose cost fell. A strategy document does not change that. One measured saving does.
A business where every job is bespoke has little routine work to remove. And one that needs a result by quarter-end should not start this quarter; mapping takes weeks, the first build takes months, and the measurement takes a quarter.
Questions people ask next
Should we hire a Chief AI Officer first?
Not necessarily. The first project needs someone with authority over one process and access to its cost figures.
Which tools should we buy?
Decide after the redesign. Most first projects need workflow automation and a language model API, and the tool you already own may be enough.
How long until we see a saving?
Plan on about six months from the start of mapping to a saving that has survived a quarter on the books.
Can we do this without outside help?
Yes, if you have someone who can map, someone who can build, and a finance team willing to baseline and verify. Few mid-sized companies have all three free at the same time.
What if our staff are afraid of being interviewed?
Keep individual conversations private and show management only organisation-level patterns. Transcript enforces this with row-level security in the database.
Lightbloom AI runs this sequence starting with Transcript: the world model first, then the visions the company chooses, built with its team. For a first look at your company, book a demo.
References
- IBM Institute for Business Value and Oxford Economics, CEOs Double Down on AI While Navigating Enterprise Hurdles (2025), https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
- European Commission, AI Act: regulatory framework for AI, Regulation (EU) 2024/1689 (2024), https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- PwC, 2026 AI Business Predictions (2026), https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html
- Stack Overflow, 2025 Developer Survey, AI section (2025), https://survey.stackoverflow.co/2025/ai
- Eurostat, Use of artificial intelligence in enterprises (2025), https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises
Keep reading.
All notes- How do you capture tribal knowledge before someone leaves? Tribal knowledge is how the company actually runs, held by the people who run it. Why handovers fail, what works, and why it is the first layer of a world model. Read →
- Why AI agents need a world model of your company AI agents stall inside companies because nobody has described how the work runs. What a company world model gives an agent, and why to build it before you deploy. Read →
- What is a company world model? A company world model is a living picture of how a company really works: every process as it runs, what hurts most, and who everything depends on. Why AI agents need it. Read →