Measuring the value
Looking up TT 38/2026 norms with one question: one code gives every material, labour and machine row — and why to ask ten items at a time
The assistant finds the code in a 10,208-code dictionary, then reads the annex page itself to pick the right column. Two neighbouring codes in one table differ by a median 14.7% in labour; one AI question is worth 24 minutes of work. How many minutes must a lookup save to pay, and what does a wrong column cost.
Norms from the book, or from the assistant: how many minutes must a lookup save to be worth its AI cost?
All it needs is the estimates per month, the coded work items in one, the share you have to look up (not known by heart), the direct cost of one estimate and an hourly cost. Lookups, today's lookup hours and AI questions follow. Everything is editable.
Derived from the figures above
Assumptions — open to see and change each one
Time
Wrong column
AI cost
This is an estimate, not a promise. Set the share looked up to 0 and all three lines are zero. An honest model has to be able to do that.
Vietnam's construction estimating norms, TT 38/2026/TT-BXD with the annexes as replaced by Official Letter 9947/BXD-VPB, are a very thick set of tables: the pages with norm tables that this tool keeps come to 1,587 pages on their own. To know what a work item consumes, an estimator opens the PDF, finds the chapter, finds the table, then finds the right column: under the same "rebar for columns and piers", which column is which bar diameter and which height. Then copies the material, labour and machine rows into the spreadsheet. A few minutes each time, and a few dozen times per estimate.
This piece asks: once a question in plain words returns the right code with every consumption component, how many minutes must a lookup save to be worth the AI call, and what does one wrong column cost?
1. What changes
In a project's assistant pane, the engineer asks in words ("non-fired brick wall under 6 m high") or by code. The assistant works in two steps:
- Search the TT 38/2026 dictionary (10,208 codes): it returns candidates with their name, column condition, unit, group and annex page. One question can look up as many as 25 work items at once.
- Read the annex page verbatim that holds the code's table before answering, to pick the right column: the names and conditions the machine split out of the tables can miss one level of condition; the original page does not.
When no candidate fits, the assistant is instructed to write «CHƯA CÓ Mû (no code yet), or to name an analogous code and say plainly that it is an analogy, for the engineer to decide.
Does the search hit. Take one real estimate, anonymised: 107 distinct codes, 83 of them in TT 38/2026. Searching with the very work names the estimator wrote, the right code is among the first 12 candidates for 69 of 83 (83%). Each search took a median of 19 milliseconds on a development machine, 39 at the slowest. The 14 misses are why step 2 exists, and why the engineer still decides.
2. What one code holds
Two neighbouring codes in the same table, Annex II page 404, both "rebar for columns and piers, height ≤ 6 m":
| Component | AF.61411 · Ø ≤ 10 mm | AF.61421 · Ø ≤ 18 mm |
|---|---|---|
| Round steel bar (kg) | 1,005 | 1,020 |
| Binding wire (kg) | 16.07 | 9.28 |
| Welding rod (kg) | — | 4.82 |
| Labour, group 2 (man-days) | 14.14 | 9.22 |
| Welder 23 kW (shifts) | — | 1.16 |
| Bar cutter-bender 5 kW (shifts) | 0.40 | 0.32 |
One digit apart in the code: labour differs by 35%, and two rows, welding rod and welder, are either there or not. Picking the wrong column does not give one wrong number; it gives a wrong set of components.
Counted across the whole dictionary: 8,453 codes have been split into consumption rows, a median of 6 rows per code, 34 at most; 5,374 codes carry materials, labour and machines all three. Across 4,271 pairs of neighbouring codes in the same table (differing only in the last digit), the labour consumption differs by a median of 14.7%; the middle half of the pairs differ by 8% to 25.8%. That is the typical price of one misread column.
3. First stream: lookup hours
How to measure it. For one week, every time the norms PDF is opened, make a tally mark and time it from opening to copying the last row.
How far to trust it: fair. A long-serving estimator knows far more codes by heart, so the share looked up varies a lot between two people in the same office.
4. To subtract: the AI cost
Every question is charged exactly its own model cost. Across the first 9 questions measured on the live system, the middle one cost 32,508 VND. At 80,000 VND an hour, that is close to 24.4 minutes of work.
From here comes the most important point of this piece: asking one code at a time is never worth it. A book lookup takes a few minutes; no lookup saves 24. Batch 10 work items into one question and the break-even drops to 2.4 minutes per item; batch 25 and it is under 1 minute. The horizontal dashed line in the chart is exactly that threshold, and it does not depend on how many lookups you do: more lookups only make the gain or the loss bigger, they cannot change its sign.
5. Second stream: fewer wrong columns
How to measure it. Take 20 codes from an estimate already submitted and check each against its annex page: count the codes with the wrong column or a missing row.
How far to trust it: low. Nobody counts how often they pick the wrong column; 3% is illustrative, and this line is the biggest in the table.
6. Adding it up
| Stream | VND/month |
|---|---|
| Lookup hours given back (5.3 hours) | 0.43 M |
| Fewer wrong columns | 3.53 M |
| AI cost (12.8 questions) | −0.42 M |
| Total | 3.54 M |
About 3.5 million VND a month, 42 million a year. But read it closely: lookup hours minus the AI cost come to only about 10 thousand VND. Nearly the whole figure comes from the wrong-column line, the least trustworthy one.
7. What this ROI stands on
The share looked up. Set it to 0: all three lines are zero. Someone who knows the norms by heart does not need this tool.
The error rate in the book, and with the assistant. Set the two equal: the biggest line is zero, and the total is just hours minus AI. The assistant can misread too; what it does differently is state each column's condition and open the original page for a person to check.
The cost of one question batching many codes. 32,508 VND is the middle of nine questions of every kind, some of which had to look at drawing images. A norm lookup only reads text, so it may be cheaper; a question batching 25 codes is longer. This cost has not been measured for norm lookups on their own.
The search does not always hit. 14 of 83 codes were not among the first 12 candidates when searched by the estimator's own wording. For those items the engineer still has to know which table to ask about again, or open the book.
8. Measuring it in your own company
- Take a baseline. For one week, record the number of norm lookups, the minutes each, and, for 20 submitted codes, how many had the wrong column.
- Run the same kind of work on the tool. Batch items into questions of 10–25 codes, record each question's cost from the credit history and the minutes spent reading the result.
- Compare like with like. Residential buildings use few unusual tables; infrastructure and equipment installation use more.
9. What to remember
The price of a lookup is not in the minutes, it is in the column. A few minutes paging a PDF are worth a few thousand dong; a wrong column is 15% of the labour of the whole work item, sometimes a whole machine too many or too few. And because every AI question has a price, batch the items and ask once. Where the unit prices put on those components come from is the subject of the piece on price sources.
One different last digit in a code is a different set of components.
The number of codes (10,208), annex pages (1,587), codes split into consumption rows, rows per code and the labour gap across 4,271 neighbouring pairs were counted on the TT 38/2026 dictionary shipped with the Xây dựng G20 tool on 31/10/2026. The two codes AF.61411 and AF.61421 were read from the original annex page. The 69 of 83 codes and the search time were measured on one real estimate, anonymised. The cost of 32,508 VND per question was measured on the live system, 29–30/10/2026. Estimates, work items, share looked up, minutes per lookup, error rates, the share that becomes real money and salary are illustrative. Replace them with your own.
Read next
-
Make-up orders from a supermarket chain: one PO in 26, and only the receipt tells
A store raises a PO and receives it the same day to pay back goods already delivered without one. The PO header still says "deliver tomorrow", so a make-up order is easily delivered twice and the goods behind it never invoiced. Measured on 1,258 POs, with a matching threshold chosen from the data.
-
One screen for a delivery crew's day: 43% of notes posted before the truck delivers, and returns nobody can record
The warehouse posts the delivery note as the truck leaves; the driver's screen read it as "done" and locked the returns field. Over 14 days, early-posted notes had a signed slip 1% of the time and a returned quantity 0.4%, against 89% and 15.5%. Reopen the note for the driver, add a vehicle & fuel tab that reads photos with AI — what that saves a month.
-
Exceptions that arrive with a reason: 82 slips to check, but only 68 need someone to think
When invoices to supermarket chains raise themselves, the remaining slips reach the accountant with a reason in words. Read by what has to be done, the reasons fall into three kinds: judgement, wait, one-time fix. Measured on 291 slips, and the price of the minutes spent searching.