Most small and mid-size businesses in Seattle, Los Angeles, San Diego, Austin and Dallas already know what generative AI can do. The gap is getting past a pilot that never ships. Here is a low-risk path that does, with the checks that stop it from stalling.
Start with a cost or revenue line, not a tool
The SMBs that succeed do not start with "let's use AI". They start with a specific, expensive workflow: support response time, quote generation, content production, lead qualification. Pick one line item, attach a number to it, and make that the success metric.
Good candidates share three traits: the task happens often, it follows a recognisable pattern, and a person can quickly tell a good output from a bad one. Poor candidates are rare, high-stakes decisions, or tasks where nobody agrees what "correct" looks like.
| Workflow | Metric to track | Risk if the AI is wrong | Good first project? |
|---|---|---|---|
| Support replies | First response time | Low, with review | Yes |
| Quote generation | Hours per quote | Medium | Yes, with an approval step |
| Content production | Pieces per week | Low to medium (brand) | Yes |
| Lead qualification | Time to first contact | Low | Yes |
| Credit or hiring decisions | Varies | High (legal, fairness) | Not as a first project |
The three-stage path
- Augment first. Put AI in front of a human: draft replies, summarise documents, suggest next steps. Low risk, fast value, and it builds trust.
- Automate the contained tasks. Once the evaluation data shows the AI is reliable on a narrow task, let it run with human approval.
- Integrate. Connect to your real systems (CRM, helpdesk, billing) so the AI takes action rather than only making suggestions.
Integration is where most of the engineering effort goes. Expect work on authentication, permissions, rate limits, error handling when the CRM or helpdesk API is down, and a log of every action the AI took so a person can undo it.
Give each stage an exit criterion. Move from augment to automate when your team has accepted most drafts with light edits over several weeks. Move from automate to integrate when the approval rate is stable and you have logs showing what the AI did and why.
Build an evaluation set before you build anything else
Collect 30 to 100 real past examples of the task: customer emails and the replies you sent, requests and the quotes you issued. Remove personal data you do not need. For each example, note what a good answer must contain. Run every version of your prompt or system against this set and score it. That gives you a number to compare instead of opinions, and it catches regressions when you change a model or a prompt.
What SMBs get wrong
- Boiling the ocean. Trying to "AI the whole business" instead of one workflow.
- No evaluation. Shipping on vibes, then losing trust on the first visible mistake.
- Over-buying. Paying for an enterprise platform when a focused build is cheaper and a better fit. See our buy-vs-build framework.
- Ignoring the data question. Sending customer records to a tool without checking the provider's data terms.
- No owner. The project belongs to "IT" or a vendor rather than the manager whose metric it moves.
Governance that fits a small business
You do not need an AI committee. You need a one-page policy covering which tools staff may use, what data may go into them, who reviews customer-facing output and how mistakes are reported. If you want a structure to borrow from, the NIST AI Risk Management Framework is free and voluntary. Also check the privacy and consumer-protection rules in the states where you and your customers operate; state-level AI and privacy law is still changing, so review it at least once a year.
Cost is no longer the blocker
With an offshore build partner, an SMB in California or Texas can often ship a real AI feature for considerably less than many onshore agency quotes, and keep iterating. When you compare proposals, compare what is included as well as the price: an evaluation set, logging and a proper handover are part of the product, not extras. Our LLM cost checklist keeps the runtime bill sane once you are live.
A 90-day plan
- Weeks 1 to 2. Pick the workflow, record the baseline, collect examples.
- Weeks 3 to 6. Build the augment version; staff use it on real work with review.
- Weeks 7 to 10. Measure against the baseline, fix the most common failures, add the approval step.
- Weeks 11 to 13. Decide: expand, integrate with your CRM or helpdesk, or stop.
FAQ
Can we just buy ChatGPT or Copilot licences?
For personal productivity, often yes, and it is a sensible first step. A custom build becomes worthwhile when the task needs your own data, a consistent output format, or a connection to systems such as your CRM.
How long before we see results?
For a contained workflow, you should see a measurable change within the first few weeks of staff using the augment version. If you do not, the workflow or the approach needs rethinking before you spend more.
Do we need to hire an AI specialist?
Not for the first project. You need an internal owner who understands the workflow and can judge output quality, plus build capacity from a contractor or partner. Hire specialists once AI work becomes continuous rather than a single project.
How Velura Labs helps SMBs
We start small and concrete: an AI Strategy & Roadmap session to pick the one workflow worth automating, then a focused build with LLM applications or AI lead generation. Talk to us about your most expensive repetitive task.