AI vs. Manual Bid Leveling

A chief estimator is asked why the desk still levels subcontractor bids by hand when a model can sort them faster. The question assumes the job is a race. Manual or AI, a level fails when the work stops at ranking cover prices. The job is to match each proposal to the drawings and the specifications, and to name who is carrying every material exclusion, before anyone trusts the low number. A faster grid finishes a bad comparison sooner.

This is for GC estimators and precon managers who have to recommend a number: on a commercial or institutional hard bid the night before it goes in, and on a CM-at-risk or negotiated GMP while the documents are still moving. The two situations ask for the same match. They punish a skipped match in different ways. How to run the pass on each delivery method is a separate piece. This one is the decision people skip before they open that guide.

There is no independent public study that shows a model levels construction bids more accurately than an estimator with a sheet. Vendor figures for hours saved are not that study. What follows is the distinction worth settling before anyone picks a method.

The grid is not the level

A bid tabulation records what each bidder submitted. A bid level asks what each bid would cost if it priced the same work. An evaluation decides who you can award. Teams already blur those three. A model that drops lump sums into columns blurs them faster. So does a senior estimator who stars three exclusions and never reads the rest of the proposal. Speed is not the test.

The test is whether each material carve-out points at a sheet, a specification section, or an allowance with a name. Tabulation, leveling, and evaluation are different jobs. A spreadsheet is a sound place to record the result of that match. Tidy columns do not perform the match. A model does not perform it because the columns filled themselves.

Spreadsheets and dedicated leveling software trade cost, habit, and an audit trail. That trade is real, and it is a different argument. Either tool can be aimed at cover prices. Either tool can be aimed at scope. The aim is the precon decision.

Hard bid: the low number becomes yours

On a design-bid-build hard bid, AIA Document A101-2017 is the familiar shape of the owner contract: a stipulated sum, a fixed price. The GC's subcontractor comparison usually happens in the last stretch before that price is locked. On private work you sometimes get a short scope conversation before award. Do not build the carry on the hope of that conversation.

Public sealed bidding is tighter. After opening, the owner's comparison is generally price and the price-related factors written into the solicitation, not a negotiation that fills missing scope. Internal leveling still has to be finished before you submit. What that file needs to show later is covered in public bid leveling. This is not legal advice. The solicitation governs.

What a price sort hides on that night is specific. The low electrical number leaves out temporary power, or fire-alarm programming, or an addendum the second bidder acknowledged and the low bidder did not. Ranked totals look like a decision. The gap is unfunded work. Submit the low with no add-back, and the gap is in your number. Carry a round figure that points at no clause, and buyout is where you will explain it. The patterns worth hunting package by package are in electrical exclusions and exclusions and qualifications.

Hard-bid leveling runs on a short clock. That is why a sorted grid gets accepted, and why a wrong low is expensive: there is often no second pricing round before the owner has your stipulated sum.

Negotiated and GMP work fail on the same skip

Negotiated work is not a hard bid with a longer clock. The commercial object is different, so a skipped scope match lands somewhere else.

AIA Document A133-2019 is written for a construction manager who advises during preconstruction and then builds the project with a guaranteed maximum price. The CM gives the owner a GMP proposal. The owner may accept it, reject it, or negotiate it. Once the owner accepts that proposal by signing the GMP amendment, the CM is bound to provide the work at or below that price. Preconstruction and construction can overlap. That overlap is the point of the form. AIA's primer on project delivery states the timing estimators already live with: when a GMP is used, the moment the construction price becomes fixed is flexible, and it varies by project. The CM is often selected on factors other than total construction cost.

CMAA's 2025 edition of Understanding Your Role describes the same sequence from the owner's side. The CM at-risk often joins during preconstruction on qualifications, with a preconstruction price, general conditions, and a fee on the table, not a fully bid trade stack. Later the owner may proceed under a GMP, a lump sum, or cost-plus, or stop and put the work out to hard bid.

The column you are tempted to sort may hold different kinds of offers:

• A competitive quote on an issued-for-bid package
• A design-assist budget at design development, with the assumptions still in the narrative
• An allowance the GMP is carrying because the documents are not there yet
• A self-perform number built from production rates, not from a proposal

Those are not one low. A design-assist budget and a hard quote answer different questions. Rank them in one column and the design-assist number looks light, or the hard quote looks high, for reasons that have nothing to do with who understood the work.

The failure mode is the opposite of bid-day panic. You can usually go back to the sub. That second conversation becomes a reason to defer the match. The sheet shows the mechanical number under the estimate. The easy reading is savings. Sometimes it is savings. Sometimes controls, commissioning, or a specification section never entered the proposal, and the delta is scope. Booking that delta as savings before the scope lines up is how a GMP inherits a bad level. The longer reconciliation, including what is not a saving, is in the leveling guide.

If the GMP amendment locks first, the gap has to live somewhere. It comes out of contingency, it becomes a quiet plug, or it returns as a change the owner treats as already inside the price they accepted. On a hard bid you meet that gap the week you submit. On a GMP you can meet it a month later, in a buyout log the owner has already seen.

AIA Document A134-2019 is the same family without a GMP: cost of the work plus a fee, with a control estimate the owner reviews as the design moves. An open book does not relax the match. It shows a bad reading of the variance to the owner sooner.

What changes between the two

The checklists for design-bid-build, CM at-risk, design-build, and negotiated work already live in the leveling guide. Use them. One grid, typed or generated, does not cover the contexts that guide separates.

What to demand from either method

Before anyone argues about whether the grid was typed or generated, require the same things:

• The rows start from this project's drawings, specifications, addenda, and your company's trade boundaries. A blank column of bidder names is a tabulation waiting to happen.
• Each material exclusion names a sheet, a spec section, or an allowance. A finding with no basis is a highlight, not a level.
• Silence stays unresolved. It is neither included nor excluded until someone decides.
• Design-assist budgets, allowances, and competitive quotes are labeled before anyone ranks them.
• The delivery is named at the top of the sheet: hard bid tonight, or GMP still open. The same delta is a carry in one case and a false saving in the other.
• A person accepts, edits, or rejects the finding. The award stays a judgment. A model that names a winner has skipped the match and gone straight to evaluation.

Where the hours go is worth measuring on your own packages. Hours do not tell you whether the output was a level. A two-hour sort and a twenty-minute sort can miss the same exclusion.

Where Piper fits

Nothing above depends on a product. A careful estimator can run this pass on a sheet. The pass gets fragile when the scope basis, the latest addendum, the proposals, and the carry live in different places, and only one person remembers how they connect.

Piper is the AI operating system for preconstruction. It keeps an evolving understanding of the project and uses that understanding to drive the work: scope, bid leveling, final review, and survey. Those are workflows of one system. Bid leveling compares subcontractors against the intended scope, not only against each other. A finding stays linked to the clause or sheet it came from. The estimator accepts, changes, or rejects it, and stays responsible for the number.

On a hard bid, that match is what you can still change before a stipulated sum locks. On a GMP, the same understanding has to survive the next design round, so a design-assist budget is not treated as a finished trade bid, and a variance against the estimate is not called savings until the scope lines up. A chatbot that has not read this project's documents, or this company's trade boundaries, will sort prices. Sorting prices is the failure, whether a person or a model does it.

Sources

• American Institute of Architects, AIA Document Synopses by Family (November 2019): A101-2017, stipulated sum; A133-2019, construction manager as constructor with a guaranteed maximum price, including the owner's right to accept, reject, or negotiate the GMP proposal; A134-2019, cost of the work plus a fee without a GMP.
• American Institute of Architects, "A primer on project delivery terms": when a GMP is used, the timing of the fixed construction price varies by project, and CM at-risk selection is often on factors other than total cost.
• Construction Management Association of America, Understanding Your Role (2025 edition): the CM at-risk often joins on qualifications, with preconstruction pricing, general conditions, and a fee, before a full trade-bid construction price.
• No independent public benchmark compared, for this piece, the accuracy of model-assisted bid leveling with manual leveling. Hours-saved claims from software vendors are not that benchmark.