Build vs Buy vs Hire AI for SMB

The three options — and why most SMBs frame this wrong

When an SMB operator decides to "do something with AI," the default mental model is almost always wrong. The question is usually framed as "should we build AI into our business?" — a yes-or-no capacity question, as if AI were a single thing you add to an org chart. It is not. AI for an operating business in 2026 is a set of specific, bounded workflows: a renewal retention scorecard, an AI phone receptionist, a commission reconciliation agent, a certificate-of-insurance generator, a silent-churn detector. Each of those workflows has a distinct make-or-buy decision.

Once you frame the question at the workflow level, three options emerge cleanly. You can build the workflow in-house by hiring engineers. You can buy a productized version of it from a vendor that already shipped the workflow to other operators in your vertical. Or you can hire a custom-build agency or consultancy to write bespoke software for you and hand over the code.

Each option has a cost structure, a time-to-value window, a maintenance profile, and an opportunity-cost shadow. Each one wins in a specific situation and loses badly in the others. The rest of this page is the framework for matching the option to the situation — with the actual dollar figures and the actual timelines, not marketing-deck versions of them.

Total cost of ownership

The sticker-price gap between these three options is an order of magnitude or more, and most operators anchor on the wrong number. Here is the honest accounting.

Build: $400K–$650K per year, minimum

A mid-level machine-learning or AI engineer in a US metro commands a base salary of roughly $170K to $210K per year in 2026, per the most recent BLS Occupational Employment Statistics release and aggregated compensation benchmarks. Fully loaded with payroll tax (roughly 10%), benefits (roughly 15 to 20%), equity or bonus (10 to 25%), laptop and software (roughly $5K), observability tooling ($10K to $20K), and cloud compute for model inference ($20K to $60K per year for a modest production workload), one engineer runs roughly $260K to $360K all-in.

You will need two. One engineer has no on-call coverage, no code review partner, no vacation backup, and no continuity if they leave. So the real floor for a functioning in-house AI capability is two engineers plus at least a part-time product owner or technical lead — roughly $520K to $720K fully loaded before you have shipped a single workflow. Subtract 15 to 25% if you are in a lower-cost metro or hiring remote in the US; add 20 to 40% for NYC or SF base rates.

Buy: $149–$2,500 per month

Productized vertical AI tools price on subscription. Ascero AI Essentials is $149 per month for two workflows installed and measured, Growth tier runs higher for full-stack deployment across a business, and per-tool Launchpad installs are priced as flat fees published on the pricing page. Comparable verticalized productized AI tools cluster in the $100 to $2,500 per month range depending on depth and breadth.

The delta is extreme. Buying Ascero AI Essentials for a full year costs $1,788 — roughly 0.3% of the annual cost of a two-engineer in-house team. That is not a rounding error; that is two different categories of spend. The subscription also includes updates as models change, which is a meaningful line item you do not have to track separately.

Hire-agency: $15K–$30K per project, plus maintenance

A competent AI build agency charges $15K to $30K for a single workflow scoped to 8 to 12 weeks of delivery. Complex integrations or highly regulated verticals push this to $40K to $75K. The price buys you the first production-ready version of the workflow and the source code to run it going forward.

Do not forget the ongoing maintenance. The underlying LLM APIs ship breaking changes on a 6-to-12-month cadence, frameworks evolve, prompt-injection defenses need updating, and your data schema drifts over time. Budget 15 to 25% of the original build cost per year to keep custom code working. A $25K build is really a $25K build plus $4K to $6K in recurring maintenance, either paid to the original agency or absorbed by someone internal.

Time to value

Cost is only half the equation. The other half is how long each option takes to produce the measurable business result that justified the spend in the first place.

Build: 4 to 6 months to first workflow

Hiring a US-based ML engineer takes 8 to 14 weeks from posting the role to signed offer, and that assumes you already know what to screen for. Add a 30-day notice period on the candidate side, a 30-day onboarding ramp, and a technical-stack decision that will eat another 2 to 4 weeks. Your first in-house engineer writes production code in month 3 or 4 at the earliest. The first actual workflow lands in production in month 5 or 6. During that entire window, you are paying full salary for zero shipped value.

Buy: 30 days to live, measured workflow

Productized tools install against your data and systems rather than being written from scratch. A Ascero AI install follows a consistent pattern: week 1 is scoping and data-access setup, week 2 is configuration to your vertical's rules, week 3 is test runs against live data, and week 4 is turning it on and baselining the metric. At day 30 you have a working workflow producing measurable output. This is the fastest path by a wide margin.

Hire-agency: 8 to 12 weeks

A scoped custom agency engagement typically runs 8 to 12 weeks from kickoff to production — roughly 2 weeks of requirements gathering, 4 to 6 weeks of build, 2 weeks of integration testing, and 1 to 2 weeks of deployment and handoff. Agencies that claim faster delivery on custom work are usually either recycling a template (in which case you are paying agency rates for what is effectively a productized tool) or skipping integration testing (in which case your maintenance burden doubles).

Maintenance burden

Every shipped AI workflow generates ongoing maintenance the day it goes live. Who does that maintenance is one of the most consequential parts of this decision.

Build: your team keeps shipping — and keeps breaking things

An in-house team is the most flexible option — they can respond to a new requirement in hours instead of weeks, they accumulate domain-specific knowledge that compounds, and they can pivot when a new model like a next-generation Claude or GPT release reshuffles the cost curve. But maintenance is the majority of their output, not the minority. Industry studies consistently find that roughly 60 to 75% of a software engineer's working hours go to maintenance, refactoring, and bug fixes rather than new-feature work. You are paying $520K+ per year for, optimistically, $150K of new-feature output.

Buy: vendor absorbs model changes, API deprecations, prompt drift

Productized tools push updates as the underlying models change. When Anthropic deprecates a model, your vendor migrates — you do not. When a prompt-injection vector is discovered, the patch rolls out to you automatically. When the LLM provider's pricing shifts, the vendor absorbs the arbitrage or passes it through transparently. This is the single most underrated benefit of the buy column, and it is invisible until you have lived through an API deprecation as an in-house team.

Hire-agency: code becomes orphan tech debt

This is where custom-agency engagements most often go wrong. You paid $25K for a working workflow in month 3. By month 15 the model has been deprecated, the original engineers at the agency have rotated onto other clients, and the code has a handful of dependencies that no one at your company understands. You now have three bad options: pay the agency a retainer to keep it running, hire internally to own it, or rebuild from scratch on the current model. Each is expensive. A custom build without an explicit maintenance plan is a future rebuild with extra steps.

Side-by-side

When building in-house actually wins

The build column wins in a narrow but important set of cases. Misreading these is what gets SMBs into six-figure hiring decisions that never pay back.

• AI is the moat, not the feature. If your AI workflow is the thing competitors cannot copy — a proprietary recommender trained on years of your closed data, a routing system that captures your specific operational edge, a model trained on your private dataset — that capability has to live inside your org. Licensing it or outsourcing it commodifies it. This is the archetypal case for build.
• You have proprietary data at a scale vendors cannot match.Ten years of transaction history, a rare clinical dataset, structured domain-specific logs nobody else has — your AI advantage is the data, and the data stays in-house for compliance, competitive, or IP reasons. You need engineers close to it.
• You are VC-backed and shipping AI as the product.If you are a Series A or later startup and AI is literally what you sell to customers, you cannot outsource the core product to a productized vendor. You have to build.
• You are past roughly $10M ARR with a data-science org.Above a certain scale the fixed cost of maintaining an in-house team amortizes across enough workflows that build becomes the rational choice. Below that line — which is the vast majority of SMBs — the math does not work.

For a 20-person insurance agency, a 40-person accounting firm, or a 12-location restaurant group, none of these conditions hold. The workflows they need are not moats; they are table stakes. Their data is valuable but not at a vendor-defeating scale. They are not shipping AI as the product. They are nowhere near $10M ARR on an AI-attached revenue line. Build is almost certainly the wrong answer for them.

When buying productized wins

The buy column wins in the largest segment by headcount — roughly every operating business from 5 employees to 100, across most verticals, for most of the workflows they actually need.

• Your problem is known and named. Silent renewal churn. Uncollected receivables. Manual COI generation. Missed phone leads after hours. Commission reconciliation. These are problems with names, and because they have names, somebody has already built a productized tool for them. If your problem fits a name, buy.
• You are sub-100 employees and want the workflow live this quarter. The 30-day install window is often the deciding factor. An SMB that needs the workflow running before the next quarter closes cannot wait 4 to 6 months on a hire or 12 weeks on an agency. Productized is the only option that fits the calendar.
• Your customizability needs fit within a config surface.Most vertical productized tools expose 30 to 80 configuration knobs. If your business logic fits inside that surface — and for 80%+ of SMB cases it does — you get all the benefits of build without any of the cost.
• You want a vendor to own the model-change treadmill.The AI stack is not finished. Models will keep getting better and cheaper, APIs will keep changing, defenses against prompt-injection will keep evolving. A productized vendor absorbs that treadmill on your behalf.

Ascero AI is built squarely for this column. The catalog spans 70+ productized workflows across 20+ verticals — insurance, accounting, legal, healthcare, restaurants, trades, RIA, real estate, e-commerce, and more — each priced transparently, installed on a 30-day window, and maintained by Ascero AI rather than the operator. The pricing page lists every tier and every per-tool install fee without a discovery call.

When hiring an agency wins

The hire-agency column wins in a specific, real, but narrower band than most agencies will admit. This is where the decision gets subtle.

• Your scope is defined but genuinely unique. You know exactly what the workflow needs to do, and no productized tool maps cleanly to it. Maybe the workflow bridges three industry-specific systems that no vertical vendor has integrated. Maybe the output format is a regulatory artifact particular to your jurisdiction. Unique plus defined equals agency territory.
• You have $25K+ to spend and want to own the code.Code ownership has real value if you plan to operate the workflow for years, extend it, integrate it with other internal systems, or eventually absorb it into an in-house team. Agency contracts deliver that ownership; productized subscriptions do not.
• You have a plan for ongoing maintenance. Either you retain the agency on a small maintenance contract, or you have internal staff who can own the code after handoff. Without a maintenance plan, the custom build becomes the orphan-tech-debt outcome described earlier. The agencies that win this column are the ones that offer a post-delivery maintenance tier; the ones that do not are a trap.
• The workflow is a one-off, not an ongoing capability.A migration, a compliance artifact, a one-time data-cleaning pipeline — these are natural agency projects. They have a beginning and an end. The maintenance treadmill matters less because the thing is not supposed to keep running indefinitely.

Notably absent from this list: "we need AI and we do not know what we want." That is the single worst reason to hire an agency, and it is the most common one. If your scope is genuinely unclear, a productized tool will force clarity faster and cheaper than a custom build will. Agencies thrive on ambiguous scope because it lets them bill more; that is a bad starting condition for the buyer.

FAQ