The case studies you read about AI — the ones with the dashboards, the percentage gains, the executive quote — are almost always written about companies with full data teams and seven-figure budgets. The version that lands on the desk of a 40-person Denver firm looks nothing like that. The model is the easy part. The expensive part is everything around it.
This post is about that everything-around-it: where small and mid-sized AI projects actually break, what the budget really looks like, and the few patterns that consistently work.
Where SMB AI projects actually fail
Three failure modes account for most of what we see, and none of them are about the algorithm.
Data the team doesn’t trust. Most SMBs have data in three or four systems that disagree with each other — QuickBooks says one revenue number, the CRM says another, the spreadsheet on the controller’s desktop says a third. Industry surveys consistently put data preparation at well over half of an AI project’s effort; in our experience with smaller shops it’s usually higher, because the cleanup work has never been done before. A model trained on dirty data produces confident wrong answers, which is worse than no model at all.
No one owns the change. A pilot that produces a useful weekly report is not a project until someone’s job changes to use it. We routinely see pilots prove value, get a round of applause, and then quietly die because the controller’s month-end close still runs the old way. The technical deployment is the cheap half; the workflow change is the expensive half.
Recurring cost gets discovered after launch. The license fee is the line item people plan for. Inference cost, retraining cost, monitoring, the contractor who reviews outputs once a quarter, the storage growth — these are not optional and they don’t show up in vendor pricing pages. Budget for them on day one.
A staged approach that actually works at SMB scale
The pattern we see succeed for businesses in the 10–200 employee range is small, sequential, and boring:
- Pick one process, one number. Not “AI for the business” — “reduce the AP-coding time on month-end close from 12 hours to 4.” If you can’t write the goal as a single number with a baseline, you’re not ready.
- Run a four-to-eight-week pilot against that one number. Use an off-the-shelf model wherever you can — custom training is rarely worth it at this scale.
- Decide explicitly what changes in someone’s job. Who runs the new workflow, who reviews the output, what happens when the model is wrong. Write it down.
- Roll out, then instrument. Log inputs, outputs, and overrides from day one so you can prove value six months later when the budget conversation comes back.
The mistake is skipping step 3. It is the difference between a project and a demo.
Where AI is paying off in SMB work right now
The wins we see most often are not glamorous, and that’s the point — they’re the ones that survive the quarter:
- Professional services (law, accounting, design): document classification, contract review triage, automated time-entry suggestions from calendar and email activity. Three- to six-month payback when the work was previously billable-but-undelivered.
- Light manufacturing and distribution: parsing supplier emails into PO updates, demand forecasting against historical sales plus a handful of external signals, automated freight-quote reconciliation. Reduces clerical headcount pressure rather than replacing it.
The pattern in both: the AI sits between two systems that previously required a human to retype data between them. That is a much easier business case than “AI for strategic decision-making.”
Watch-outs before you sign
- Audit and explainability. If the output drives a financial entry, a compliance filing, or a customer-facing price, you need a logged reason for every answer. Many of the cheaper AI products do not provide this.
- Vendor lock-in. Tightly tuning a prompt or workflow to one model is fine for a pilot, dangerous for a system you’ll run for three years. Keep the integration thin enough to switch.
- Data residency and confidentiality. Decide what the model is allowed to see before the contract, not after. For regulated work (HIPAA, PCI, SOC 2 customers) this is non-negotiable.
- The honest twelve-month picture. A useful first project at SMB scale typically runs 4–8 weeks of implementation, then six to nine months of iteration before it is doing what the original pitch promised. Plan for the iteration, not just the launch.
Next step
If you are looking at an AI project and trying to figure out whether the case study you saw last week applies to your actual business, get in touch. We do realistic scoping for SMBs in the Denver area and tell you when the answer is “not yet.”