Most mid-sized employers have tried something with AI by now. A pilot in the finance team, a chatbot on the website, a tool the sales group picked up on its own. The harder question is whether any of it is paying back, and whether you could prove it to your owner or board.
In our experience the answer is usually unclear, and the reason is rarely the software. It is that nobody decided what "paying back" meant before the project started.
There is no shared yardstick
A plant that wants fewer unplanned stoppages, a controller who wants a faster month-end close and a sales leader who wants cleaner pipeline data are all chasing different results. All three can be getting a good return. None of them can be compared to the others, or to a competitor you read about.
So stop looking for an industry benchmark. Build your own, and build it first.
Define the number before you buy the tool
Before any project starts, write down four things on one page:
- The specific workflow being changed, named at the task level
- The current baseline: hours, error rate, cycle time or cost per unit
- The target and the date you will check it
- The person who owns the result, by name
If you cannot fill in the baseline, you are not ready to start. Spend two weeks measuring the work as it is today. Many teams skip this, and then six months later cannot say whether anything improved.
Hours saved is a fair starting measure, but it is not the end of the story. Saved time only becomes money when it is redeployed, when overtime falls, when a hire is avoided or when throughput rises. Decide which of those you expect.
Where projects actually get stuck
The sticking points tend to be operational, not technical:
- Fitting into existing systems. The tool works in a demo and fails when it has to read data from your ERP, your CRM or a spreadsheet someone built in 2016.
- Accuracy and compliance. Someone has to check the output, and someone has to decide what the tool is not allowed to touch. Customer data, payroll and regulated records need rules before use, not after.
- Building and maintaining it. A prototype is one thing. A tool that is monitored, fixed and updated is another, and it needs a named owner.
- Data quality. If your item masters, customer records or job codes are inconsistent, the AI will be inconsistent too.
Each of these is a people problem as much as a technology one. They are solved by someone with the time and skill to do the work.
The roles that decide the outcome
Mid-sized companies rarely need a large AI department. They need a few specific people in the right places.
- A business owner for each use case. Usually an operations manager, controller or sales leader who cares about the result and has authority over the process.
- A data or analytics person who can clean, connect and monitor the data the tool depends on. This is often the missing hire.
- An integration or engineering resource, permanent or contract, to get the tool working with your systems.
- A reviewer with enough subject knowledge to catch bad output, such as an experienced accountant, estimator or quality lead.
Contract talent fits well here. A 6 to 12 month engagement for integration or data cleanup lets you prove a use case before you commit to a permanent headcount. If the work proves out, you will know exactly which role to hire for.
Train the people who already work for you
Tools sit unused when supervisors and team leads do not understand how they fit into the day. Budget time for training, and let managers set the expectation that the tool is part of the job, not an optional extra. Adoption follows what managers actually ask for.
A short set of working rules helps too: what the tool may be used for, what must be reviewed by a person, and who to call when it gives a strange answer. One page is enough to start.
Start small and measurable
Pick one or two workflows where the baseline is easy to measure and the risk is low. Invoice coding, quote preparation, maintenance scheduling, first-pass resume or order-entry checks are common starting points. Run each for 60 to 90 days, compare to the baseline and then decide whether to expand, adjust or stop.
Stopping a project that is not working is a legitimate result. It frees the budget and the people for something better.
What to do next
Take your current AI efforts and test each against the four-line page above. Where the baseline, target or owner is missing, fix that before spending more. Then look at the roles on that list and mark which ones you do not have in-house today.
If the gap is a data person, an integration resource or an operations lead who can own the rollout, a Traverse recruiter can help you find them, on contract or permanent.