Where AI adds value in project controls

0:00 Where AI actually earns its keep

If you listen to vendor pitches, AI is about to run your entire project controls function. If you listen to sceptics, it's a toy that makes things up. Both are wrong. The value in twenty twenty-six is real, but it's uneven. It's strongest where there's lots of structured history, repetitive text work or pattern-finding across large datasets. It's weakest where data is poor, or where judgement and accountability are central. In this lecture you'll get a practical map of where AI helps in controls, a simple scoring template to choose your first pilot, and two worked examples: narrative drafting and anomaly detection. Plus a watch-me-do-it on prompting for a status report. By the end, you'll be able to pick a use case that's valuable, safe and easy to verify.

0:56 Why it matters

Why does it matter? Remember the control cycle. Most teams spend far too much time collecting data and drafting reports, and far too little on analysis and action. AI's best contribution is shifting that balance: drafting the first version of a narrative, checking a schedule for logic problems, flagging odd cost postings, so people can spend their time judging and advising. But there's a trap. If your first use case is too ambitious, say, a fully automated forecast that drives commercial decisions, it'll either fail visibly or succeed invisibly and dangerously. Either way, trust collapses. The first use case sets the tone for everything after it.

1:42 The concept: a use-case map

Here's the map. For schedules, AI can detect open ends, odd lags and unrealistic durations, and predictive models can estimate which activities are likely to slip. The planner confirms. For cost, models can produce an independent EAC and flag unusual forecasts, and outlier detection can find miscoded costs, duplicate invoices and sudden progress jumps. The cost controller and analyst decide. For risk, language models with retrieval can suggest risks from past projects and lessons learned. Owners validate. For documents, they can extract obligations, notice periods and change clauses from contracts. Commercial and legal verify. Computer vision can estimate installed quantities from photos, drones or BIM models. And generative AI can draft narratives and answer natural-language questions over project data. Notice the pattern: in every row, there's a human role. That's not a nice-to-have. It's the design.

2:41 Choosing a pilot: the 25-point score

So how do you choose? Score each candidate from one to five on five criteria. Value: time saved, earlier warning or better decisions. Data readiness: is there history, is it clean, can you access it? Risk if wrong, where lower risk scores higher. Ease of human verification: can a person quickly check the output? And fit with your current tools and skills. The rule I suggest: proceed if the total is eighteen or more and nothing scores a one. Narrative drafting, schedule quality checks and anomaly flags usually score well: valuable, low risk and easy to check. A fully automated forecast feeding a claim scores badly on risk and verification, however exciting it sounds. Here's the key idea: pick the use case you can verify, not the one that impresses.

3:38 Worked example one: narrative drafting

Let's look at a pilot that worked. A fictional infrastructure owner in the UK produced twenty-five monthly project reports, and controls leads spent about two days each drafting narratives. The pilot gave an approved AI assistant access to the structured EVM tables, the change log and the risk register for each project. Not raw emails. It drafted narratives in the standard template. Controls leads reviewed, corrected and signed off. What errors did review find? Two typical ones: attributing a variance to the wrong cause, and confident wording where the data was ambiguous. With a checklist-based review, drafting time fell substantially while quality held. And every draft and every human edit was kept on record. That's a safe, valuable, verifiable first pilot.

4:31 Worked example two: anomaly detection in Dubai

Now anomaly detection. A fictional contractor in Dubai ran an outlier model over cost postings every week. It flagged three things. Invoices posted to control accounts that showed no progress. A supplier invoice for exactly the same amount as one already paid. And a sudden forty-point jump in percent complete on a package with flat labour hours. When the analysts investigated, two were genuine errors: a miscoding and a duplicate. The third was valid: a prefabricated module had been delivered, which legitimately earns a lot of progress with few site hours. So one in three flags was a false alarm. Is that a failure? Not at all. The value was focusing human attention on three items out of thousands, not replacing the humans who knew about the prefab delivery.

5:27 Watch me do it: a status-report prompt

Let me show you the prompt I use for a first-draft status report, in whichever enterprise assistant your organisation has approved, whether that's Microsoft 365 Copilot, ChatGPT Enterprise, Claude or something built into your controls platform. I start with the role: you're supporting a project controls analyst. Then the data: I paste the structured table, never confidential commercial documents unless the tool is approved for them. Then the template: headline, cost, schedule, risks, decisions required. Then the rules. Use only the data provided. Cite the control account and metric for every claim, in brackets. Separate facts from interpretations. List what's uncertain. Don't invent causes; write 'cause to be confirmed'. What comes back is a draft I can check in minutes, because every number points at its source.

6:22 What AI doesn't do well, and common mistakes

Let's be clear about limits. AI doesn't understand causation from correlation alone, so it can give plausible but wrong explanations. It doesn't know facts that aren't in its data: site conditions, relationships, informal agreements. It can't take accountability: a signed forecast is a professional commitment by a person. And it struggles with rare, novel events that have little historical precedent. The common mistakes follow. Starting with the most ambitious use case. Deploying without a baseline measurement of how long things took before, so you can't prove value. No verification step. And assuming AI output is neutral, when it reflects its training data and your data. And one rule to never break: don't paste confidential commercial data into tools your organisation hasn't approved.

7:15 Recap and try this now

Let's recap. AI in project controls is valuable where there's structured history, repetitive text work or pattern-finding at scale, and where a person can verify the output quickly. Use the twenty-five point score to choose: value, data readiness, risk if wrong, ease of verification and fit. Start with narratives, schedule quality checks or anomaly flags, not automated forecasts. Feed tools structured data, ask for cited outputs that separate facts from interpretation, and keep human sign-off and a record of edits. Your try-this-now: score three potential AI use cases for your team using the template, and pick one pilot. Then write down how long that task takes today, so you can measure the difference.