AI for Project Scheduling

Project scheduling is building and maintaining the programme itself. The activities, the durations, the logic that connects them, the critical path that falls out of it, and the float that tells you where you have room. This is distinct from writing about the programme, which is its own discipline. Here the output is the plan. There the output is the prose. The reason that distinction matters for AI is simple. Scheduling is a calculation. A language model does not calculate, it predicts what an answer looks like, and a predicted critical path is worthless. WORKED EXAMPLE: WHERE THE FLOAT ACTUALLY IS A fit-out sequence on a single floor. Seven activities, finish-to-start logic throughout. | Ref | Activity | Duration (days) | Follows | |-----|----------|----------------|---------| | A | Partitions | 10 | - | | B | First fix M&E | 15 | A | | C | Plaster and dry lining | 12 | B | | D | Ceiling grid | 8 | C | | E | Second fix M&E | 10 | D | | F | Floor finishes | 6 | C | | G | Decoration | 8 | E and F | Two paths run through this from C to G. Path via ceilings: A + B + C + D + E + G = 10 + 15 + 12 + 8 + 10 + 8 = 63 days Path via floors: A + B + C + F + G = 10 + 15 + 12 + 6 + 8 = 51 days Project duration is 63 days, and the critical path runs through the ceiling grid and second fix. Floor finishes carry 63 - 51 = 12 days of total float. Check it the other way: F runs in parallel with D and E, which together take 18 days. F takes 6. So F has 18 - 6 = 12 days of float. Same answer, which is what you want from a check. Now the useful part. A five-day delay to floor finishes costs you nothing at all. A five-day delay to the ceiling grid moves completion to day 68, one for one. Two activities, similar duration, similar cost, completely different consequence. Ask a site manager which of those two they would chase and most will say floor finishes, because it is visible and it is behind. The programme says otherwise. One more thing this example shows. If floor finishes slip by 15 days rather than 5, they consume the 12 days of float and become critical themselves, and the project now has two critical paths. Float is not a permanent property of an activity. It is a relationship between paths, and it moves. WHY AI CANNOT DO THIS PART Everything above came from adding durations along paths and comparing them. That is a forward and backward pass, and it is what scheduling software does deterministically and correctly every time. A language model asked "which of these activities is critical" will give you a fluent answer derived from which activity sounds important. On the example above, a plausible wrong answer is that first fix M&E is critical because it is the longest activity. It is on the critical path, but so is everything before G on that branch, and length has nothing to do with it. Use scheduling software for the calculation. P6, Asta, MS Project, whatever your business runs. Then use AI for the parts that are language. WHERE AI IS GENUINELY USEFUL IN SCHEDULING Building the first activity list. Give it a project type and scope and it will produce a sensible list of activities and a reasonable sequence to react to. Faster than a blank page, and the value is in what you delete. Sense-checking logic in plain English. Describe your sequence and ask what is missing. It will ask about commissioning, about the lead-in on long-lead items, about whether you have allowed for testing. Not because it understands your job, but because it has seen the shape of many programmes. Turning a programme into readable text for people who cannot read a bar chart. Most clients cannot, whatever they say. Drafting the assumptions and exclusions register that should sit alongside every programme and usually does not. Interrogating durations. Ask it to list every activity where the duration looks optimistic for the quantity involved. It will be wrong sometimes and it will still surface two you had not questioned. Producing the lookahead. Pulling a three or four week window out of the programme and turning it into something a site team will actually read, with the activities, the resources needed and the things likely to block them. The programme holds the data; the lookahead is a communication job, and that is language work rather than calculation. It is one of the few scheduling outputs AI can genuinely draft, because the hard part has already been done by the software. WHEN NOT TO USE IT Any calculation. Critical path, float, dates, resource levelling. All of it belongs in the software. Delay analysis. Establishing what caused a delay and whether it was critical is a technical exercise on real programme data, and the recognised methods exist precisely because the intuitive answer is often wrong. Anything going into a claim. The programme underpinning a claim needs to survive expert scrutiny. FREQUENTLY ASKED QUESTIONS What is the difference between total float and free float? Total float is how long an activity can slip without delaying project completion. Free float is how long it can slip without delaying the next activity. They are frequently different, and the difference is what causes arguments. An activity with 12 days of total float but zero free float can slip a single day without moving the end date, and still push the activity immediately after it. Reporting only total float hides that. Who owns the float? Contractually this is one of the most argued questions in construction scheduling, and the answer depends on your contract and its amendments rather than on any general principle. Some forms and amendments address it explicitly, many do not, and the position on whether an employer delay may consume contractor float is genuinely contested. Read your contract, and if float ownership is likely to matter on your job, take advice before you need it rather than during the argument. Can AI build a programme for us? It can draft an activity list and a plausible sequence, which is a real time-saver at the start. It cannot produce a programme, because a programme is a calculated network with dates and float, and those come from software operating on logic you have set. Treat AI output as a checklist to react to, then build the actual programme properly. How detailed should a construction programme be? Detailed enough to manage and coarse enough to maintain. The failure mode at both extremes is the same: the programme stops being updated. A tender programme with 60 activities and a construction programme with 600 are both appropriate in their place. The test is whether you can genuinely update it every period. A 2,000-line programme that gets updated properly twice a year is worth less than a 300-line one that is current. What makes a programme defensible if things go wrong? That it was accepted, that it was updated regularly with actual progress recorded, and that the logic is sound rather than propped up with constraints. Programmes riddled with hard date constraints do not behave like networks, so delay does not propagate and the critical path becomes meaningless. Record actual start and finish dates as you go. A programme that was updated monthly during the job is worth vastly more, later, than one reconstructed afterwards.