November 2026 · 8 min read
Beijing Put AI Consulting in the Vendor Catalog

Key Definitions
Provider pool The list each region compiles on attachment 2 and submits to the ministry by 1 December 2026; the national pool is assembled from those lists and published in due course. Targets are 2,000 firms by end-2026 and 3,000 by end-2027.
Provider archive The attachment 1 form, with four sections: basic information, service capability, service cases, undertaking letter. Attachment 3 sets the gate: at least three staff directly engaged in the four service categories with social insurance records, and 10 cases backed by contracts for firms under three years old.
Forward deployed engineers (FDE) The policy asks providers to build FDE teams rooted at the customer site. The buying form shows what that means: engineers placed inside the business, running requirement study through acceptance on site.
Two documents landed in August that describe the same market in two languages that do not translate into each other.
The first is a notice from the General Office of China’s Ministry of Industry and Information Technology, issued 27 August 2026 and published on 31 August. Its first service category for an “AI application service provider” is consulting and planning, defined as business strategy, industry analysis, management systems, and project implementation design. The second is a procurement notice from the Jiangsu branch of China Tower for AI application support services. It buys engineers on site: the text specifies an FDE deep on-site model, staff placed inside the business, and a one-year term. The budget is 138,000 yuan before tax, which is about 19,000 dollars at the announcement’s own 6 percent VAT line, for twelve months.
One document defines who qualifies. The other defines what a day costs. The price schedule has no line for consulting.
What the notice actually wrote
The notice carries the document number 工信厅科函〔2026〕414 and runs with five attachments. Its opening definition is precise: a provider is an enterprise or institution that supplies consulting and planning, delivery and implementation, operations management, and security governance around a client’s AI requirements. The four categories appear in that order, with consulting first.
Attachment 1 breaks each category into services. Consulting and planning covers AI-related business strategy, industry analysis, management system design, and project implementation planning. Delivery and implementation covers platform and tool operation, custom application development, and systems integration. Operations management covers terminal maintenance, system monitoring and repair, software updates, and product iteration. Security governance covers protection, risk warning, and hardening. A fifth category, “supporting services,” covers training and testing and evaluation.
One detail in that list gets skipped in most commentary. All four consulting items are verbs. None of them requires the output to be a document. The policy takes a position on who thinks the work through and writes the implementation plan, and it takes no position on the format. That matters for what follows, because it means the policy did not exclude reports. The procurement layer is what failed to give a report a unit of account.
The targets are counts. By the end of 2026 the national provider pool should exceed 2,000 firms; by the end of 2027 it should hold at least 3,000; provinces that host a national AI application innovation pilot zone should hold at least 100 locally by the end of 2027. The reporting schedule is specific: each region submits its pool information to the ministry’s science and technology department by 1 December 2026 and updates it periodically; each province submits at least 10 service teams, and each team has one lead provider plus at least two upstream or downstream participants; a summary report is due by 1 December 2027.
The sentence quoted most often sits in the fourth task: the notice encourages providers to build forward-deployed engineer (FDE) teams, rooted in the customer’s site, to secure deployment. The People’s Daily overseas edition of 4 September summarised it the same way, as support for training bases and for developing FDEs to deepen the talent pool. Both texts place FDEs in the workforce section rather than in the procurement section.
The notice also closes a door that the market has been trying to open. Its organisational support clause says that regions which implement the programme well will receive preference in relevant policies and projects. The preference attaches to regions, not to firms. Nowhere does the text promise that a listed provider wins priority in purchasing. “Being in the pool means being on the government’s approved vendor list” has no official source, and it should stop appearing in sales conversations.
The gate has a shape, and the shape matters
Attachment 3 is the reporting requirement, five items, each with named evidence. Read it closely, because this is the first time a national document has defined an AI service provider with materials that can be checked rather than described.
The first item is basic eligibility: a legally registered entity in China whose main business covers AI services; no serious credit violations for the firm, its legal representative or its controller in the past three years, evidenced by a Credit China query; no third-party intellectual property infringement, evidenced by an undertaking letter. The second covers qualifications: third-party testing and evaluation requires CMA or CNAS certificates; public websites, mobile apps and mini-programs require ICP filing and related pre-approvals; generative AI services with public opinion or social mobilisation attributes require registration with the cyberspace administration. The third is the team: in principle at least three people directly engaged in consulting, delivery, operations or security governance, evidenced by social insurance payment records. The fourth is capability: mastery of relevant technology and application patterns, the ability to identify a client’s core need, premises, equipment and tools, and a complete, usable service manual or standard documentation. The fifth is case volume: at least 30 cases in the past three years for firms with three or more years of experience, at least 10 for firms with less, evidenced by service contracts or application certificates.
Three numbers get quoted: three people, ten or thirty cases, one manual. Together they draw a line between a small team that can deliver and an intermediary that owns a slide deck. A two-person shop fails item three, because the evidence is a social insurance record rather than a business card. A firm two years old with twenty projects behind it can fail item five just as easily, if the contracts and acceptance records live on someone’s laptop, because every row in the table needs a stamped certificate attached.
The second-order effect is larger than the threshold. For the past two years, small teams took work project by project, finished, and moved on. Contracts were signed on the client’s template, and acceptance often amounted to a confirming email. The archive now asks for those scattered actions to be re-laid as ten checkable rows, each with a contract value and a delivery period. Work performed without retained evidence counts as work not performed, on this form. It is an asset inventory exercise, not a writing exercise.
The template also shows what the reader of the form will look at. Each service description is capped at 200 characters, the company profile at 500, each case at 500 with a contract or application certificate attached. Submission is a stamped scan plus an editable file, and the undertaking letter needs the legal representative’s signature and the company seal. The table reads contract values, delivery periods, client names and stamps. It does not read how long the methodology is.
The buyer has already published a price: 138,000 yuan a year
China Tower’s Jiangsu branch ran an inquiry for AI application support services for 2026 to 2027. The notice went out on 10 August 2026, the documents were available from 10 to 14 August, and responses closed at 09:30 on 19 August. The numbers are quotable as published: an estimated budget of 138,000 yuan before tax, 146,280 yuan including 6 percent VAT, a maximum response price of 138,000 yuan before tax with bids above it rejected, one winning supplier taking 100 percent of the scope, a one-year service and contract term, staff on site within 20 calendar days of signing, no consortia, and no subcontracting or transfer.
The scope paragraph is the important part. It specifies an FDE deep on-site model, dedicated technical staff placed inside the business, embedded in the engineering workflow, working from live scenarios, raw operational data and existing reporting systems to find automation needs, delivering AI reporting agents and office automation assistants, and completing requirement study, solution design, development, validation, iteration and acceptance on site throughout. The phrase used for the commercial arrangement is “AI technical labour on-site service,” applied to data processing, reporting automation and process digitalisation support.
That paragraph fixes the unit of account. The budget pays for technical labour on site. The scope is arranged by business process. Acceptance happens on site. Solution design appears in the text, but as one phase of the resident engineer’s work rather than as a separately sold artefact.
Divide the numbers and the unit price appears. One resident engineer against 138,000 yuan is 138,000 yuan per person-year. Three engineers make it 46,000 each. The tax-inclusive 146,280 yuan spread over twelve months is 12,190 yuan per month. The price ceiling equals the budget, so there is no room to bid higher. What a bidder can compete on is how to staff the work inside a fixed envelope. There is also a track-record threshold: since 1 August 2023, at least 100,000 yuan including tax of cumulative work in comparable categories, namely systems integration, data processing, AI application development or data governance. To bid on a 138,000 yuan contract, a firm first proves it has already delivered something near that size in the same category.
So the price anchor in this market is a resident headcount-year, not a consulting day rate. Inside a 138,000 yuan annual budget there is no row to allocate to diagnosis, strategy or a method. At ordinary consulting day rates, the same budget buys a couple of weeks. This is the actual position of report-style consulting here: it was not rejected, it has no budget line.
The same delivery shape landed abroad in the same quarter
On 10 November IBM and Microsoft announced Imagine Enterprise. IBM’s newsroom states that IBM Consulting brings more than 6,000 pre-built AI agents and assets built on IBM Consulting Advantage, that Microsoft supplies the AI, data, cloud and security foundation, and that forward-deployed engineers from both companies plus GitHub support qualifying engagements moving from experimentation to production. The first industry frameworks are financial services, insurance and retail. Mohamad Ali, senior vice president and head of IBM Consulting, put the demand side plainly: clients no longer want AI that helps people do the same work a little faster, they want to build an agentic business.
On 3 November, the European private equity firm Hg announced an agreement to invest, with participation from Anthropic and angel investors, in a new AI transformation company. The announcement describes an independent entity launching later this year, built on Claude, with forward deployed engineers, aimed at moving clients beyond pilots to measurable operational impact. Terms were not disclosed.
Three unrelated parties, a Chinese ministry, an IT services firm with a platform partner, and a European fund with a model lab, described the delivery shape in one sentence within the same quarter: engineers go on site, assets get reused. The Chinese version is policy language encouraging FDE teams. IBM’s version is product language, 6,000 agents and assets plus forward-deployed engineers. Hg’s version is investment language, forward deployed engineers on Claude. None of the three mentions reports.
Evidence grades differ and should be stated. The 6,000 figure is IBM’s own count and has not been independently verified. The Hg announcement gives a shape and a goal without headcount, price or client numbers. Both are used here to show that the delivery shape is converging, not to size a market.
Where a report can still sit
Put the three documents on one page and the space for a report becomes narrow. The archive has three columns: service type, service case, service team. None of them is “delivered document.” The case column requires a contract or an application certificate, and a report usually has no acceptance record to attach. The procurement notice pays for labour and elapsed time, and a report is not on the work list.
Two positions do exist. The first is inside project implementation planning: when the plan is bound to delivery, it can be written into the service description as the opening of on-site work, within the 200-character cap. The second is inside testing and evaluation, which the policy places in supporting services rather than in the four core categories. Both positions share a property. The value of the report has to be recognised by something else, or paid for by something else.
That yields a conclusion worth carrying out of the article. In China, report-style consulting is becoming a sunk cost incurred before procurement. It happens before the bid and before the project is approved, to decide whether the work should exist and how acceptance will be defined. What buyers pay for it usually sits in another budget line. A firm that wants to enter the pool on the strength of reports alone runs into a structural fact: every cell in that form asks for a contract value and a delivery period.
The one seat priced per delivery: an evaluation
On 19 August 2026 the Chongqing Development and Reform Commission opened a competitive selection for a contractor to evaluate the effects of AI applications in the city’s tendering and bidding system. The legal basis cited is the State Council’s opinion on deepening the “AI plus” initiative (国发〔2025〕11号) and a 2026 implementation opinion on AI in tendering and bidding (发改法规〔2026〕195号). The work covers the full chain of AI deployment in legally required construction tenders, with emphasis on bid document screening, assisted evaluation and collusion detection, assessed quantitatively across five dimensions: economic benefit, compliance and risk control, operational efficiency, social benefit, and technical security. The funding line states a maximum price of 200,000 yuan covering the entire scope with no additional payment, and completion within three months of signing.
Divide this against the China Tower contract and the difference is the unit. Two hundred thousand yuan over three months is about 66,700 yuan a month. The Tower contract’s tax-inclusive 146,280 yuan over twelve months is 12,190 yuan a month, roughly one fifth. The more important difference is that the Chongqing seat does not require anyone to live on site. It buys one defined evaluation.
It carries its own gate, and the gate has a different character. Participants must be registered in China with independent legal personality and the right to sign contracts; must have comparable project experience; must field a project lead who holds a senior professional title at the level of associate professor or above, or an equivalent rank, and who personally organises, guides and participates in the work; must show no bad contract performance in three years; must describe quality assurance and control measures. The submission requires a research application form, five paper copies stamped on the first and last pages at the entity’s official seal, one electronic copy on a disc, sealed and delivered to Room 716, No. 1 Hengming Road, Tiangongdian Street, Liangjiang New Area, Chongqing, by 18:00 on 31 August 2026.
That window has closed, so this is not a lead anyone can follow today. Its value in this article is as a format sample. It shows a buyer willing to pay for one evaluation rather than for elapsed time, and it shows a seat that is held by a titled individual. Add the policy’s placement of testing and evaluation in supporting services, plus the CMA or CNAS requirement for third-party evaluation, and access to that seat is not a matter of writing a good report.
Three things that can start today
For the person writing a procurement document: turn the acceptance section into data items. State the definition, the baseline and the recomputation method, and who compares what at which point. Do not stop at “submit one evaluation report,” because that sentence cannot be accepted. Split the quote into two parts as well: labour by person-month or person-year, assets by package, with the evaluation set, the acceptance-criteria library and the scenario templates listed separately. The same 138,000 yuan then shows whether the money bought elapsed time or something reusable.
For a firm preparing a provider archive: start with an inventory. List every project from the past three years that can produce a contract and an acceptance record, lay them out in the attachment 3 format, replace missing contracts with copies and missing acceptances with confirmation letters. Then write the service manual, which is item four and the only part of the archive that shows a method. The service team column can carry the FDE roster; the policy’s fourth task explicitly encourages it and it costs nothing to include.
For us: we walked our own materials against attachment 3 and the blockers were the social insurance headcount and the verifiable case count, not the writing. So we did not spend this cycle competing for a place in the catalog. We put the effort into acceptance criteria and evaluation sets, which serve two purposes at once, as the standard a client accepts against and as the asset an archive can show.
Our reading, and what would prove it wrong
Over the next year, AI consulting in China will split into two purchasable objects: on-site service priced by labour and elapsed time, and evaluation priced per delivery. The strategic report in between has no budget line and will exist as a step before those two, funded from somewhere else. This is not a mood. It follows from the unit of account in three documents rather than from what any of them claims.
The reading has a checkable implication. When provincial pool lists become public, look at the service types providers tick. If far more tick delivery and implementation than tick consulting and planning, the system gave consulting a name and the market gave it no price. The lists are public and anyone can check. If instead the consulting tick rate is high and per-engagement consulting lines start appearing in provincial procurement, the reading is wrong and should be revised.
Boundaries belong here too. The notice gives no date for publishing the national pool, only that it will happen in due course. The 138,000 yuan and 200,000 yuan figures come from two different buyers, one for a year on site and one for three months of evaluation, and they cannot be combined into a single firm’s pricing power; they are used here as evidence about units of account. The 6,000 figure is IBM’s own, and the Hg announcement disclosed neither terms nor headcount. The Chongqing window closed on 31 August. Company-level filing windows are set locally, region by region, and this article did not check them province by province; it uses only the 1 December provincial reporting deadline.
References
All figures were read from the following pages on 11 November 2026, each returning HTTP 200 without redirect. Policy quotations come from the full text and attachment 3; procurement amounts come from the notice bodies.
People’s Daily overseas edition, 4 September 2026, report on the AI application service provider programme (four tasks, including training bases and FDE development)
Full text and five attachments of the MIIT notice (document 工信厅科函〔2026〕414, 27 August 2026; provider definition, 2,000 and 3,000 firm targets, 1 December reporting, the three-person and 10/30-case thresholds in attachment 3, the five service categories and archive layout in attachment 1)
Phoenix Finance, 31 August 2026 (independent confirmation of document number and publication)
Heilongjiang provincial forwarding notice, 11 September 2026 (how province and city levels organised the survey)
IBM Newsroom, 10 November 2026: IBM and Microsoft announce a new AI agent platform (6,000-plus agents and assets, forward-deployed engineers, first industries, Ali’s statement)
PYMNTS, 10 November 2026 (independent coverage of the same announcement)
Hg announcement via RNS, 3 November 2026 (independent entity, launch later this year, built on Claude, forward deployed engineers, measurable operational impact)
China Tower Jiangsu branch inquiry for AI application support services, 10 August 2026 (138,000 yuan before tax and 146,280 including 6 percent VAT, 138,000 ceiling, FDE on-site scope text, 100,000 yuan track record)
Chongqing Development and Reform Commission selection notice, 19 August 2026 (200,000 yuan maximum, three months, five evaluation dimensions, project lead title requirement, 31 August deadline)
FAQ
How does the 414 notice define a provider+
An enterprise or institution supplying consulting and planning, delivery and implementation, operations management and security governance around a client’s AI requirements. Consulting is listed first; a fifth supporting category covers training and testing and evaluation.
What are the hard requirements for the pool+
At least three staff directly engaged in the four categories with social insurance records; 10 cases for firms under three years old or 30 for older ones, each backed by a contract or certificate; a service manual or standard documentation; and CMA/CNAS, ICP filing or cyberspace registration where the relevant service is offered.
Does being in the pool mean approved vendor status+
No. The notice promises preference in policies and projects to regions that implement the programme well. It never promises listed firms priority in purchasing, and that claim has no official source.
Is the 138,000 yuan a consulting fee+
It buys resident labour. The Jiangsu notice specifies an FDE on-site model with staff placed inside the business for one year, 146,280 yuan including 6 percent VAT, capped at 138,000 yuan before tax. No line in that budget covers diagnosis or strategy.
Can a firm still bid the 200,000 yuan evaluation+
No, submissions closed at 18:00 on 31 August 2026. The gate is also narrow: comparable experience, a project lead at associate-professor rank or above who participates personally, and five stamped paper copies delivered sealed.
Is there still room for report-style consulting+
Two positions, both attached to something else: project implementation planning inside the consulting category, as the opening of on-site work, and testing and evaluation inside supporting services. A report cannot be a case row, a service category, or a unit of price on its own.
Run the same check on your own execution log
Send a redacted 30-day agent execution log and get an eight-row table in 48 hours: one row per technical item, each cell carrying its numerator, denominator and criterion, gaps sorted by whether a schema change is needed. First three engagements are $1,200.
Related Articles
Enterprise agents converge on a single front door
Google folded chat, coding and scheduled work into one agent with one memory graph. 4 questions decide whether that memory and its skills can leave.
Only 13% of CIOs see AI value: walk-away rules
Gartner asked 2,002 CIOs: 13% see significant value. Firms with walk-away rules are 2.3x and 2.6x likelier to lead, yet only 31% and 17% have them.
38 percent AI growth inside a 2 percent budget
AI spend in a service budget rose 38 percent while the envelope rose 2 percent. In a fixed envelope the cap lands on the most successful workflow.
AI pays inside functions and stalls at the seam
Surveys of 6,000 executives and 1,719 firms agree: AI lifts output inside functions while nine in ten report no productivity gain. The seam is unmeasured.