How do I choose an automation partner?
Judge them on how they estimate, not on what they have built. Anyone can show you a demo. The firms that finish on budget are the ones who watched the process before quoting, can name the exceptions they expect to find, and will tell you when not to build.
Four questions separate them quickly: how did you arrive at this number and who did you watch; whose accounts will this run in; what happens when it breaks at 2am; and what would you tell us not to automate. A firm that cannot answer the fourth is selling, not advising.
Price is a weak signal in both directions. A quote well below the rest usually means the exceptions, the testing and the handover were left out, and those are the parts that make it survive.
First decide what shape of help you need
Picking the wrong category costs more than picking the wrong firm within it.
| You need | Right choice | Why |
|---|---|---|
| One small, well-specified thing | Build it yourself, or a freelancer | A two-step connection between two modern apps is an afternoon. Paying a firm for it is paying for overhead you do not need. |
| Several processes, unclear which matter | A firm, starting with proper scoping | The diagnosis is the hard part. Buying builds before you know the priority is how budget disappears. |
| Continuous availability, constant small changes | A hire | That is a role, not a project. No firm will be as responsive as someone sitting with you. |
| A twelve-month pipeline and someone to own it | A hire, plus consulting to set standards | An underused full-time automation capability either costs you money or quietly becomes general IT. |
In-house vs agency vs freelancer works the economics through properly.
The four questions
1. "How did you arrive at this estimate, and who did you watch?"
The single most predictive question. An estimate produced from a one-hour call with a manager has priced the process as the manager believes it works, which is reliably not how it works. The person who actually runs it knows about the customer whose orders are handled differently, the spreadsheet that gets reconciled on Fridays, and the approval that is technically required but always skipped.
A good answer describes shadowing: who they sat with, for how long, and what surprised them. If nobody watched, the number is a guess, and an overrun is likely.
2. "Whose accounts will this run in?"
It should be yours, your platform subscription, your credentials, your vault. Then ask the follow-up that actually matters: if I revoke your access tomorrow, does anything stop working? The answer should be no. If revoking a contractor's access takes down production, you have a dependency nobody disclosed. What happens if we stop working together covers the patterns.
3. "What happens when this breaks at 2am?"
Not if. Credentials expire, vendors deprecate endpoints, someone renames a column. The answer you want describes three things: how it fails (it should stop, not continue with wrong data), who gets alerted and how, and what the underlying records look like afterwards. A firm that has not designed the failure path has not finished the design. See what happens when an automation breaks.
4. "What would you tell us not to automate?"
The question that sorts advisers from order-takers. Every real business has processes that should not be automated: too rare to pay back, still changing shape every month, or high-stakes with no signal when they go wrong. A firm that has looked properly will have found at least one and will say so unprompted. A firm that says everything you mentioned is a great candidate is selling, not advising.
"How will you measure whether it is right?" The answer must involve an evaluation set built from your historical cases with known correct answers, not "we tested it and it looked good". Without that, nobody can tell you the error rate, which means you will learn it from customers.
Warning signs
- A quote before anyone watched the work. Covered above, and the most important.
- Percentage-of-savings pricing. It sounds aligned and it is not. It creates an incentive to overstate the baseline, and the baseline is the thing you cannot independently verify.
- Reselling the platform. If a firm takes a margin or a commission on the tool it recommends, the recommendation is not advice. Ask directly whether they hold reseller agreements.
- "100% automated." Nothing reaches 100%. A realistic target for document processing is 75% to 90% straight through after tuning, with the rest going to a review queue. Anyone promising the whole thing has not counted the exceptions.
- Headcount-reduction framing. Mostly false at SMB scale, and it tells you they are selling to a board rather than solving a problem.
- No mention of a parallel run. Going straight from testing to live skips the step that catches the cases nobody mentioned, while the manual process is still there to catch them.
- Case studies without numbers, or with implausible ones. Ask what the baseline measurement was and who took it.
- Documentation as a line item. If handover costs extra, the default is a system only they can maintain.
Signs you have found a good one
- They ask what happens today when something goes wrong, before asking what you want built.
- They talk about exceptions early and specifically, because exceptions are most of the engineering.
- They tell you the full running cost, not just the build cost: their own monthly fee and what goes to vendors.
- They are willing to lose the sale, recommending an off-the-shelf product, or telling you to wait until a process settles.
- The price is fixed before work starts, and the scope that price covers is written down.
- They plan a parallel run and will not skip it to hit a date.
Comparing two quotes fairly
Quotes are rarely comparable as written, because the cheap one is usually cheap by omission. Normalize them on six things before looking at the totals.
- Exception handling. How many branches of "except when" are in scope, and what happens to the rest?
- The review queue. Anything with a confidence threshold needs an interface for the cases that fail it. A spreadsheet of flagged items is not one, and people abandon it within a fortnight.
- Testing against real data. Against your actual historical cases, or against examples they made up?
- Parallel running. Included, and for how long?
- Documentation and handover. Included, or extra?
- Who owns the accounts. Yours, or theirs.
Once both quotes cover the same six, the gap usually closes considerably, and where it does not, you can see what you are paying for. If you want a second view on a proposal from another vendor, we will give you one in writing and it costs nothing. Sometimes the answer is that the quote is fair and you should proceed with them.
Want a second opinion on a quote you have received?
Send us the proposal. We look at scope, assumptions, platform choice and price against what the work should realistically involve, and give you a written view. It costs nothing, and sometimes it says go with them.
You will talk to Fayez or Kareem. Nobody else.