Notes ·
AI agents: build vs buy vs managed
An honest map of the five categories competing for the SMB AI-agent budget, with pricing bands and where each one is the wrong choice.
Owner-led firms in the 5–50 person range have a specific version of the AI-agent problem that the general market coverage doesn’t address. They are too small to hire a full-time AI or ML engineer. They are too complex, and too regulated in some verticals, to trust a no-code template. They have workflows that are worth automating but not worth burning a founder’s month rebuilding the fifth time it silently breaks.
Five categories of vendor and tool currently compete for that budget. Each solves a real problem for some segment of the market. None of them is the right answer for every firm. This piece is an honest map of the five, with pricing, and — most importantly — where each is the wrong choice.
The five categories
1. DIY builders and course-sellers
Gumroad courses, cohort programs, and YouTube channels teaching owners how to spin up their own AI-consultancy practice or build agents in-house on Zapier / Make / n8n / OpenAI. The pitch is empowerment: “you can build this yourself in a weekend.” Pricing runs from $0 to about $2,000 for a course; ongoing costs are the tool subscriptions plus the owner’s time.
The DIY category works when the workflow is genuinely simple, the owner has time and technical curiosity, and the failure blast radius is low. It fails when the workflow touches customer data or regulated processes, when the owner discovers a course of study is not the same as production experience, or when the “weekend build” starts silently breaking three weeks in and there’s no one to call.
2. Engineering-heavy dev shops (often offshore)
Custom-build shops selling AI-flavored software development. Named examples in the mid-market: Execution Point, Layer3Labs, Stratryx, Flow Digital, and a long tail of boutique agencies with AI practices grafted onto broader software offerings (characterizations based on public materials as of 2026-08). Pricing bands: $10,000 to $50,000+ per project, with hourly rates typically $150–$300/hr.
Dev shops are the right answer when a firm has a bounded, well-specified custom build — for example, an integration between two systems where no off-the-shelf option exists — and the firm is ready to own the software afterward. They are the wrong answer for ongoing operations. Most dev shops are structured to ship a project, invoice, and move on; the code lands in your account and the firm inherits both the ownership and the maintenance burden. If the workflow needs to keep running for two years, the shop’s incentives and your operational reality are not aligned.
3. Managed AI operators
The category Phronimos plays in. Firms that don’t just build the agent — they run it. Named examples: Phronimos, Phronesis Labs, phron.ai, System Prompt AI, and a handful of newer entrants positioning around “AgentOps” as a service (again, based on public materials as of 2026-08). Pricing bands: $999 for a diagnostic, $3,500+ for a scoped implementation, $5,000/month and up for managed operations.
Managed operators are the right answer for a firm that wants outcomes rather than code — that would rather pay a monthly fee to have someone accountable for the agent working next month than a one-time build fee and inherit an unmaintained system. They are the wrong answer for a firm that wants to own the code and staff the ongoing work themselves.
4. Platforms (Zapier / n8n / Make with AI features added)
The horizontal automation platforms have all shipped AI features — Zapier’s AI Actions, Make’s OpenAI modules, n8n’s LangChain nodes. Pricing bands: $20 to $500/month per platform, occasionally higher at enterprise tiers.
Platforms are the right answer for a single low-stakes automation — inbound-lead classification, calendar reminders, straightforward form-to-CRM pushes. They are the wrong answer for a workflow with more than four steps, for anything with a legal or reliability audit trail requirement, or for anything where “it silently stopped running for two weeks” is a real cost. The platforms genuinely try to make monitoring easy, but the failure modes documented at /notes/honest-failure-modes-of-business-ai-agents/ apply in full: a Zap that reports success and delivers empty output is a Zap that will keep reporting success and delivering empty output until a human notices the downstream consequence.
5. Fractional CTOs
Individual practitioners or small partnerships selling part-time technical leadership. Pricing bands: $8,000 to $25,000/month for meaningful engagement (roughly 10–40 hours/week).
Fractional CTOs are the right answer when the firm needs broad technical leadership across engineering, hiring, roadmap, and vendor selection — not specifically AI operations. They are the wrong answer when the need is narrow and operational rather than strategic. A fractional CTO is expensive because their scope is broad; buying a fractional CTO to keep one agent running is buying a fire truck to water the lawn.
Where Phronimos is the wrong choice
The most useful part of a comparison is where the vendor tells you not to buy them. Three cases where Phronimos is the wrong fit:
- A 100+ person firm with an in-house engineering team. Hire an AI or ML engineer directly. A managed operator makes sense when the firm doesn’t have the engineering density to run agents itself; once that density exists, in-house ownership is faster and cheaper.
- A firm that needs a one-shot custom software build — not ongoing operations. If the goal is to specify, build, ship, and own a piece of custom software, a dev shop or a contract engineer is the right vendor. Phronimos’s economics only make sense when the workflow keeps running and someone stays accountable for it.
- A firm that only needs one low-stakes automation. If the workflow is a single-step form-to-CRM push and it doesn’t matter if it stops for a week, use Zapier or Make directly. Paying a Reliability Review fee to check a five-step workflow is proportionate; paying it to check a single Zap is not.
The honest ICP for managed operators
The 5–50 person owner-led firm has a specific profile that makes the managed-operator category work economically. All of the following tend to be true:
- The owner or a very small operating team makes the buying decision.
- The firm runs on a mix of vertical-specific software (AMS for insurance, PMS for law, CRM for real estate) and horizontal tools (Google Workspace or Microsoft 365, Slack or Teams, one or two automation platforms).
- Workflows are painful but not exotic — client intake, document generation, follow-up scheduling, invoice chasing, referral tracking.
- The firm has never hired an engineer and doesn’t intend to.
- Regulatory context is real (NAIC bulletins for insurance, ABA guidance for law, RESPA for real estate) but the firm doesn’t have compliance staff.
At the bottom of that range (5–10 people), the firm often can’t yet justify a $5,000/month retainer and buys the Review or the Sprint first. Toward the top (30–50 people), a firm is close to the point where hiring in-house starts to make sense — and the honest advice may be to plan for that transition rather than sign a multi-year retainer.
Concrete pricing bands, side by side
| Category | Front door | Ongoing | Who owns the code |
|---|---|---|---|
| DIY builders | $0–$2,000 course | Tool subs + owner’s time | Owner |
| Dev shops | $10k–$50k+ project | Owner runs it, or re-hire the shop | Owner |
| Managed operators (e.g. Phronimos) | $999 diagnostic | $5,000/month and up | Operator runs, owner licenses |
| Platforms | $20–$500/month | Same | Owner |
| Fractional CTOs | $8k–$25k/month | Same | Firm |
Prices for the managed-operator column are Phronimos’s /pricing/. Prices for the other categories are typical bands, not offers — check the specific vendor. The dev-shop and fractional-CTO bands especially vary widely by geography and specialization.
How to choose in practice
Three questions, in order:
- Is the workflow worth automating at all? If the annualized cost of the manual process (volume × time-per-instance × loaded hourly rate) is under $5,000/year, the answer is probably to leave it manual and revisit later. A calculator that runs this math lives at /tools/cost-of-manual-process/.
- Do you want to own the code, or the outcome? If you want to own the code, you’re looking at dev shops or DIY. If you want to own the outcome and pay someone else to keep it running, you’re looking at managed operators or (at larger scale) fractional CTOs.
- Does the failure mode matter? If the workflow silently breaks and no one catches it for two weeks, what does that cost the firm? If the answer is “nothing much,” a platform is proportionate. If the answer is “a customer, a compliance finding, or a lawsuit,” the failure modes at /notes/honest-failure-modes-of-business-ai-agents/ are the ones to plan against — and platforms don’t detect three of the four.
The Phronimos position, stated plainly
Phronimos is a managed AI operator. The offer stack — Review at $999, Sprint from $3,500, Fractional AI Officer at $5,000/month — is on /pricing/. The reliability standard behind the retainer, including detection targets and SLA credit terms, is at /reliability/. The four failure modes the Review checks against are at /notes/honest-failure-modes-of-business-ai-agents/.
If a reader has finished this piece and concluded Phronimos is not the right choice for them, this piece has done its job. The category is real, the pricing bands are real, and the right vendor depends on the shape of the firm — not the shape of the pitch.
Last reviewed: 2026-08-13.