Tracking government tenders: manual sweeps, keyword alerts and AI monitoring

Ask a team how they track public tenders and the first thing you hear is usually a worry about missing one. Press further and it turns out that what gets missed is rarely the notice itself. The notices are public, searchable and sitting there for anyone to read. What gets missed is the window in which a decision could still have been made — by the time somebody finally reads the thing properly, there is no longer enough time to prepare.

There are roughly three ways to do this. What separates them is not how automated they are. It is where each one loses things.

Approach one: the manual sweep

Assign someone to check the notices on a fixed rhythm, daily or weekly, and pull out whatever looks relevant.

This approach is underrated. If the categories you care about are narrow, the volume is steady, and the person genuinely understands your eligibility position, a manual sweep produces the best judgement of the three. They read the whole notice, not a summary and not the line a keyword happened to match.

It fails in three ways, and none of them make a sound.

First, it is attached to one person. They take leave, they resign, or they have a bad week, and the tracking simply stops. Nobody notices, because a missed notice was never going to appear on a list in the first place. The absence leaves no evidence.

Second, under pressure it degrades silently. The quality of the sweep drops from reading to skimming, then to skipping this week, and at no point does anything raise an alarm.

Third, it keeps no record. You cannot answer what you passed over last month or why, so you cannot review the judgement either. A year in, nobody can say whether the standard in that person’s head is still the standard they started with.

Approach two: keyword alerts

Define a set of keywords and get notified when a notice matches. The channel might be a subscription feature, an email alert, or a scraper somebody wrote. The form varies; the mechanism is the same one: rule-triggered push.

It is steadier than a manual sweep because it does not get tired. It also has two structural limits.

First, keywords match strings, not meaning. The same underlying requirement gets written up by different agencies as outsourced system maintenance, ICT equipment servicing, or custom system development. Set two of those and the third quietly does not exist for you — and nothing tells you that a notice failed to match the words you chose.

Second, and more commonly: an alert can tell you a notice exists, but not whether you qualify for it. Eligibility rules, track-record thresholds, bid bonds, delivery windows — all of that is inside the notice, while the alert usually carries no more than a title. So somebody still has to open it, read it and judge it. The work has not been reduced. It has only been moved to a different starting point.

The noise does a second kind of damage. When eight alerts in ten are irrelevant, whoever receives them starts archiving without reading by about week three. At that point the system reports itself as working normally while the actual result has fallen back to where it was before — and nobody files a report about it.

Approach three: AI monitoring

Parse the notice and its eligibility criteria, filter out what is impossible against hard rules, then produce a fit assessment and gap indication for whatever survives, notify only above a threshold you set, and keep tracking deadlines and amendments.

The difference from keyword alerts is what arrives. One says a notice matched your string. The other says you meet the criteria here but lack one track record item, or you do not qualify for this one because of a specific condition. The second message is enough to decide whether to spend someone’s afternoon on it. The first is not.

It is worth being equally clear about what this does not do. AI monitoring addresses which opportunities are worth evaluating. It does not submit bids and it does not download or complete tender documents for you. The bid decision and the paperwork stay with your team. The closest honest description is a first-pass screener that never gets tired and will explain why it filtered something out.

It has its own way of losing things, which is the next section.

The real difference: the shape of what gets missed

All three miss things. The question when choosing is not which one misses least, but which kind of miss you can live with.

• Manual sweeps miss on attention. What gets lost is whatever that person had no time for, did not read, or misjudged that day. The misses are random and cannot be reconstructed afterwards.
• Keyword alerts miss on vocabulary. What gets lost is whatever was written in words you did not anticipate. These misses are systematic — the same category goes missing repeatedly, and quietly.
• AI monitoring misses on stated criteria. It executes the rules you gave it, so if those rules are wrong, too narrow, or never properly articulated, it will faithfully filter out the right opportunity.

The third kind has a property the others lack: it can be inspected. You can look at what was filtered out and against which rule, then change the rule. What the first two lose leaves no trace, so there is nothing to correct.

This is also why the real barrier to adopting AI monitoring is not technical. It is whether you can write your hard exclusion criteria down. Which categories you never touch, below what scale you do not bid, which missing track record is always fatal, how short a preparation window makes it an automatic pass. Teams that cannot write this down will not get faster with any tool, because the judgement was never being made explicitly in the first place. It was distributed across a few people’s instincts.

Choosing

You do not have to jump to the end.

If your categories are narrow, the volume is steady and you have someone who understands the eligibility position, a manual sweep is enough. Write the criteria down anyway, so it stops depending on one person.

If the categories are widening, or the person tracking them has another job as well, keyword alerts are a reasonable next step. Accept the noise, treat them as a way to narrow the field, and do not mistake them for a screening result.

When the team’s time is going into a loop of opening notices, reading them and discovering they do not qualify, that is the signal to change the method. The problem to solve is no longer knowing that an opportunity exists. It is deciding whether it is worth reading to the end.

What we are building

TenderRadar is PRESIRE’s answer to the third approach. It is currently in internal deployment and pilot use, and it is not open as a public multi-tenant service. It does not submit bids and it does not complete tender documents.

If your organization tracks procurement continuously and you recognize that loop eating your assessment capacity, start with an actual opportunity. Tell us about the one you missed last month, or the one that took far too long to rule out. That is more useful than talking about features.