Buying an ATS for an agency, start to finish

Ask what happens to your data if you cancel. Specifically: what formats the export produces, whether it includes attachments and notes with their original authorship and timestamps, how long you have to retrieve it after the contract ends, and whether the export is self-service or a paid service request. This one question is more predictive of a good relationship than any feature answer, because it tells you whether the vendor's commercial model depends on you being stuck.

Ask which of the features you were shown are generally available today and which are in build. Ask for that in writing. A roadmap is not a defect; a roadmap presented in the present tense is.

Ask about the seat minimum, the renewal price mechanism and whether the contract carries an annual escalator. Ask what happens to the price if you shrink from eight desks to five in a downturn. Ask whether the CRM capability you were shown is on the tier you are being quoted, because it frequently is not.

A trial that everybody enjoys and that ends in a warm feeling has told you nothing. Design it so that a specific failure is possible and would be visible.

Pick two recruiters, not the enthusiastic one and not the one who has been complaining since 2019. Give them a real slice of data, including the awkward records. Set three tasks that must complete inside the trial: source and submit a candidate against a live requisition end to end, import a hundred records and reconcile the count, and produce the report you actually show clients or the board.

Write down beforehand what would make you say no. If nothing on that list would make you say no, you are not evaluating, you are shopping. Then hold a short session at the end where the two recruiters describe what they had to leave the system to do; that list is the real gap analysis.

Object types migrate at very different quality levels, and the difference is not obvious until afterwards. Candidate and client records with standard fields move cleanly. Custom fields move if somebody maps them by hand. Notes usually move as text but frequently lose their author and sometimes lose their timestamp, which is exactly the metadata that makes a note useful in a dispute.

Attachments are the ones that bite. They are often stored separately from the records, they are large, and export limits mean partial transfers that look complete. Activity streams, email associations and consent history commonly do not survive at all, because they were never modelled the same way in both systems.

The defence is arithmetic. Count everything in the source before you start: records per object, attachments per record, notes per record, oldest and newest timestamps. Run a test load of a representative slice. Count the same things on the other side. Investigate every discrepancy rather than accepting a plausible explanation, because the plausible explanation is usually wrong in an interesting way.

Resistance to a new system is rarely irrational. A billing consultant has a workflow that produces their commission, and you are proposing to disrupt it during a quarter they are judged on. Treating that as obstruction guarantees a slow, expensive roll-out.

The things that work are unglamorous. Move the commission calculation into the new system early, because whichever system pays people is the system of record regardless of policy. Make the first two weeks about capture only, with no new reporting demands, so the desk experiences the tool as a place to put things rather than a surveillance layer. Pick one respected consultant, give them a genuine say in configuration, and let them tell the room.

Set a hard date for the old system going read-only and announce it at the start. Open-ended parallel running is how firms end up paying two licences for a year and getting the benefit of neither.

Firms that have never had structured process treat implementation as the moment to encode every rule they have ever wished for. Nineteen stages, mandatory fields at each, approval steps for things nobody has ever approved. The desk works around all of it within a month, and the data becomes worse than the spreadsheet it replaced because now it is confidently wrong.

The safer approach is to configure the minimum that supports the questions you already ask, then add. A stage earns its place when its absence stops you answering something real. A mandatory field earns its place when a missing value has actually cost money.

Keep a list of the things you decided not to configure. Six months in, most of them will look unnecessary, and the two that do not will be obvious and easy to justify.

Agree the evidence before go-live, while you are still capable of being disappointed. Adoption first: what share of submittals were logged the same day, how many candidate records created this quarter carry a consent basis, how many desks produced their weekly numbers from the system rather than a spreadsheet.

Then one or two operational measures with the definitions written down. Submittal-to-interview ratio by client is a good one because it changes behaviour: a client who interviews one in nine of your submittals is a briefing problem, and now it is visible.

Resist grading on time-to-fill alone. It moves for reasons that have nothing to do with your software, and unless you separate client-side delay from agency-side delay you will credit or blame the system for something it did not do.