We wrote this guide for the buyer who has been asked, by their CEO or by their CTO or by themselves, whether on-demand engineering is worth procuring, and from whom. The guide is deliberately vendor-neutral. We name the questions we’d ask any vendor, including ourselves, and we explain how to read the answers. If the conclusion you reach after reading is to choose a competitor, we’d rather you do that than choose us for the wrong reasons.
The right question is not which vendor. The right question is what shape of help, in which moments, with what guarantees.
Step one: name the moments you’re buying for
On-demand engineering is, at root, an answer to specific moments. Before you talk to vendors, write down the moments. The most common, in our experience:
The internal-builder moment. A non-engineer in your company has built something with an AI tool and is about to ship it. The migration moment.A team is moving systems and needs senior engineering capacity that doesn’t exist on payroll. The integration moment. The product-engineering team has built a thing that works internally but cannot survive contact with external systems. The off-hours moment. Production is on fire and the on-call rotation is thin.
Step two: ask about the bench
The bench is the set of engineers the vendor will route to your moments. Ask:
How big is the bench, and what’s the bar? A small bench at a high bar is usually better than a large bench at a mixed bar. Tier-based vendors will tell you the bench is large; ask how many of them are at the seniority you need.
How is the bench compensated?1099 contractors paid per-press will, on average, optimize for closing the press quickly. W-2 employees on retainer will, on average, optimize for the customer outcome. Both shapes can work. Ask which one you’re buying.
Who routes the press, and on what criteria? A vendor that lets the customer pick the engineer sounds appealing and produces a worse outcome at scale (the most-requested engineer gets buried; quality drops; no fairness). A vendor that routes algorithmically, by stack and availability, scales better.
What’s the geographic distribution?If you need 24/7 coverage, the only honest answer is engineers in three time zones, working their local-day shifts. Anything else is a graveyard premium hidden in the price, or it’s a fiction.
Step three: ask about the press
How does the vendor define the unit of work?
What’s the median time-to-engineer?Anything over five minutes is not on-demand; it’s scheduled. Anything under sixty seconds with a flat-priced offer is too good to be true and the math should be inspected. Honest numbers are between 60 and 180 seconds.
What’s recorded, and where does the recording live? For regulated industries, the answer matters. For everyone else, it still matters: the session record is the audit trail.
What’s the SLA when the press doesn’t resolve?Most vendors don’t have one. Ask anyway. The answer tells you whether the vendor is taking ownership of the outcome or the activity.
Step four: ask about pricing
Three pricing shapes are common. Each has tradeoffs.
Per-press flat fee.Honest for short presses; expensive for long ones. Predictable for budgeting. Watch for vendors that cap session length and re-charge for “continuation.”
Pre-purchased minutes / hours. Honest economically; complicates internal accounting if usage spikes. Watch for expiration policies that effectively are use-it-or-lose-it.
Monthly retainer with rollover. Cheapest at sustained volume; underutilized when volume is uneven. Watch for retainer sizes the vendor pushes you to that exceed observed press volume.
Step five: ask about scope discipline
The most common vendor failure is scope drift. A press that was supposed to be twenty minutes turns into a four-hour engineering project that nobody priced for. Ask:
What’s the time-box on a press, and what happens at the box? A clean answer: the engineer says they’ve hit the box, names what’s left, and the customer explicitly chooses to extend or close. A bad answer: the engineer keeps going. A bad answer disguised as a good one: the engineer wraps and a follow-up is silently scoped.
Can the vendor say no?A vendor that has never refused a press is a vendor that ships things they shouldn’t. Our own refusal rate sits around 2-4% in regulated environments. If a vendor claims 100% resolution, ask for the methodology.
Step six: pilot before you commit
Run a four-week pilot before signing for twelve months. The pilot should include:
One-to-three of your hardest expected presses.Don’t pilot with the easy ones; you’re evaluating ceiling, not floor. Multiple time zones if you need 24/7. Schedule a press at 3am local; verify the answer arrives without quality loss. An integration that touches a real system. Vendors are good in sandboxes. The integration cliff is where most weak vendors fall over. A failure-mode press.Set up a press where the right answer is “don’t ship.” Watch how the vendor handles the conversation.
Step seven: the procurement traps
Three patterns we see kill otherwise-good buys.
Buying for a moment that doesn’t recur. A vendor you use four times in the first quarter and never again is a vendor you should not have signed annually with. Start month- to-month if uncertain.
Mistaking the demo for the bench. The engineer who runs the sales demo is, at most vendors, not the engineer who picks up the press. Ask to meet two engineers from the bench; the difference will be informative.
Skipping the BAA / DPA review. If your environment has regulated data, the security review is the procurement critical-path, not the commercial one. Start it on day one.
What “good” looks like, in a sentence
A software engineer arrives in seconds, reads the build, says what holds and what doesn’t, makes the smallest correct change, narrates it, hands the build back, and leaves. Anything more than that is overhead. Anything less and the vendor isn’t the buyer’s asset; it’s the buyer’s liability.