Software can now prepare a production, purchasing, or quotation decision in seconds, yet still leave the business waiting. For aerospace and defence suppliers, the return depends on deciding which actions an agent may complete, which need judgement, and what a correctly completed transaction actually costs.
Andy Demir ·21 October 2026 · 15 min read
Executive Summary
Enterprise AI agents are moving from analysis into the workflows that turn information into commitments: production planning, parts procurement, quotations and supplier documentation. The agents SAP presented at Hannover Messe in April 2026 prepare production recommendations that planners still approve, while AAR's Airvoyant platform, launched the same month for airlines and MROs, ranks supplier quotes and is presented as the foundation for autonomous ordering. For an aerospace or defence supplier, the open question is no longer whether software can prepare the work. It is which actions the software may perform within configuration control, export control, and contractual limits, and what that changes in the economics of the whole transaction. Faster preparation alone can simply move the queue to the approver. The useful measures are end-to-end cycle time, cost per correctly completed and authorised transaction, the age and difficulty of exceptions, and whether the gain appears as cash, capacity or customer response. My reading is that over the next twelve months credible deployments in the sector will combine agent interpretation, rules-based enforcement and targeted human approval rather than broad autonomy. The decision worth making now is a bounded, measured trial within a single repeatable workflow, with explicit criteria for widening or narrowing delegation.
In This Essay
- A production recommendation still has a decision boundary
- Permission to enter a system is only the beginning
- Faster preparation can move the queue
- Exceptions decide the economics
- The strongest case for keeping people involved
- What credible deployments may look like by October 2027
- A trial that gives management a usable investment decision
- Questions for the C-Suite and the Board
- References
A production recommendation still has a decision boundary
At Hannover Messe in April 2026, SAP described a Production Planning and Operations Agent that lets planners release production orders in natural language while the software checks material availability, capacity and scheduling constraints. It can propose alternative components or rescheduling options, which planners then review and approve. SAP gave the second quarter of 2026 as the planned general availability. The announcement describes intended capability, not independently measured customer outcomes. Its design still makes the central point of this essay visible: the agent's ability to prepare a decision and the planner's authority to release it are kept apart (Metzger, 2026).
In aerospace and defence, that boundary carries more weight than in most industries, because the information inside a recommendation has consequences outside the application that displays it. In a production system certified to AS9100 or EN 9100, an alternative component is more than a planning choice; it falls under configuration control and may require design authority or customer approval before it can be fitted. A rescheduled work order can shift a delivery commitment to a prime contractor, and a purchase requisition becomes a contractual obligation once it becomes an order. These are illustrative paths rather than claims about any particular product or manufacturer, and they explain why the business event the software completes matters more than how fluent its conversation looks.
The evidence available in October 2026 is public and uneven: vendor announcements, standards work and surveys of mostly US and multinational respondents, none of which measures realised returns at a European aerospace or defence supplier. Read with that limit in mind, it still points to a practical investment question. How much additional value becomes available when a correct recommendation can proceed to an authorised action without another avoidable hand-off, and what does the organisation have to define before that is safe?
AI agents are software systems that pursue a goal through a sequence of steps, using connected tools and responding to intermediate results. In industrial work, they can interpret documents, assemble evidence and resolve routine variations before passing a transaction to existing controls. The argument here is that their return depends on the economics of that entire path. Intelligence, permission, data quality and responsibility all have to survive the step from preparation to action, and the supplier that defines that step clearly is likely to capture more of the value than the one that simply buys the more capable model.
Permission to enter a system is only the beginning
On 29 September 2026, the National Cybersecurity Centre of Excellence at the US National Institute of Standards and Technology published a summary of the comments it had received on its concept paper on software and AI agent identity and authorisation. More than 600 commenters from industry, government and academia responded, and the centre chose the software development lifecycle, through its DevSecOps project, as the first implementation use case for showing how agents can be identified, authenticated and authorised. This is standards development rather than a finished industrial standard, but it shows that the conditions under which software may act are becoming an engineering problem in their own right (NIST NCCoE, 2026).
Survey evidence points the same way. In Deloitte's 2026 State of AI in the Enterprise research, 21% of 3,235 business and IT leaders across 24 countries who were directly involved in their organisations' AI programmes said they had a mature governance model for agentic AI. Deloitte's own description of maturity includes clear boundaries on which decisions an agent may take alone and which need human approval, monitoring of agent behaviour, and audit trails of agent actions. The figure measures what respondents report, not audited control effectiveness, yet it places the delegation question at the centre of governance rather than at its edge (Bayiates, 2026).
The commercial distinction is easy to state. Access to a supplier database does not confer the authority to approve a supplier, and permission to draft a quotation does not establish which price, delivery promise or contractual condition the business has authorised. A technically valid system operation can therefore sit outside the intended business mandate. In the aviation aftermarket, the gap is concrete: a part bought quickly from an unverified source is of little use if it arrives without the release documentation that allows it to be fitted. In defence supply, where technical data may be controlled under ITAR or the EU dual-use regulation, the question of which identity may open a file, and from which jurisdiction, is already a licensing matter that an agent with broad read access simply inherits.
That is why recent aftermarket launches deserve a close reading. In April 2026, AAR launched Airvoyant, a procurement platform for airlines and MROs, whose agents analyse inbound supplier quotes and generate purchase recommendations based on historical transactions, pricing patterns, and supplier performance. AAR presents the platform as laying the foundation for autonomous ordering and says that agents for demand consolidation, vendor optimisation, and automated negotiation are expected later in 2026 (AAR Corp., 2026). The direction is clear, and the capability may well prove valuable. The question for any airline, MRO or supplier connected to such a platform is which of those steps its own delegation of authority allows the software to complete.
For an illustrative purchasing workflow, a usable mandate might permit an agent to prepare an order against an approved supplier and an existing contract within a defined budget, while reserving a new bank account, an unapproved source or changed contractual terms for separate review. Boundaries of this kind make routine work easier to complete, because the agent does not have to rediscover the limits on every transaction. The control earns its cost when it removes uncertainty from the ordinary path while keeping consequential variations visible to the people accountable for them.
Faster preparation can move the queue
Microsoft's 2026 Work Trend Index, published in May, surveyed 20,000 AI-using knowledge workers across ten markets, including Germany, France, Italy, the Netherlands, and the United Kingdom, in a field study conducted by Edelman Data x Intelligence between February and April 2026. It placed 19% of those AI users in the group where individual capability and organisational readiness reinforced each other. Both measures are self-reported composites, so the finding describes perceived readiness rather than realised productivity, and it does not show that any particular approval structure causes delay. As a vendor publication, it is best read as a directional signal (Microsoft, 2026).
It is still a useful prompt to examine the process surrounding the user. A commercial team can prepare more quotations without increasing its capacity to review additional exceptions, and procurement can assemble supplier files faster while the approval meeting maintains its existing rhythm. When upstream work accelerates, and downstream capacity stays constant, more work arrives at the next decision point. Whether a queue then grows depends on arrival volumes, reviewer capacity and the variety of cases, so faster software alone does not settle the outcome. The question is sharper in aerospace and defence because volumes are rising. Deloitte's 2026 outlook for the sector expects supply-chain pressure to persist through at least 2027, as defence primes increase output of missiles, munitions, and drones, and aircraft manufacturers push for higher production rates, stressing every tier of the supplier base (Berckman et al., 2025). For a tier-two or tier-three supplier, that means more requests for quotation, more first-article and conformity documentation and more change notices arriving at the same small group of reviewers.
An illustrative quotation process shows the investment problem. If an agent assembles technical and commercial information more quickly, the team may reach the customer sooner. If every quotation still enters the same review queue, the release date may barely move. The gain can still be worthwhile, especially if preparation quality improves, but it is a different gain from the shorter customer response time assumed in many business cases.
Different functions also work to different clocks for legitimate reasons. Finance evaluates commitments, programme teams protect delivery, quality examines conformity and commercial teams respond to the customer. Delegation needs agreement across those clocks. A sensible design makes routine cases explicit enough to proceed, gives every exception an owner and supplies reviewers with the evidence they need. Counting approved transactions, their elapsed time and their corrections then shows whether that agreement improves the business outcome.
Exceptions decide the economics
KPMG's US AI Quarterly Pulse for the third quarter of 2026, based on 314 C-suite and business leaders at US organisations with annual revenue of US$1 billion or more surveyed between 24 July and 25 August, reported that 62% were building, developing or deploying AI agents, up from 53% the previous quarter, and that 49% had defined high-risk use cases in which autonomous decision-making by agents is not allowed. The first figure combines development stages and should not be read as a production adoption rate. KPMG also reported that the share of organisations building controls into the agents themselves had fallen to 30% from 43% two quarters earlier, as many relied on guardrails already in place (KPMG LLP, 2026).
Those findings describe adoption alongside boundaries. The next economic question is what happens to the cases that cross a boundary. An agent may save time on ordinary transactions while referring people to ambiguous records, conflicting instructions, or missing evidence. If each referral arrives with a clear explanation and the relevant documents, specialist attention is well used. If the reviewer has to reconstruct the entire case, the apparent savings at the preparation stage reappear as verification work.
For a defined workflow, a useful primary measure is cost per correctly completed, authorised transaction. That cost includes software consumption, integration, human review, exception handling and correction. It can be accompanied by the share of transactions completed without unplanned intervention, provided that share has a stable denominator. Comparing a pilot containing only straightforward cases with a baseline containing the full mix of transactions would say little about the likely result at scale.
The exception rate alone is insufficient. Several small referrals may take less time than one unusual case involving engineering, quality, and commercial colleagues at once, so the age of the exception queue and the time needed to resolve its hardest cases should be reported alongside the average. This points to a second-order effect that is easy to miss: as routine work is automated, the remaining human workload may become more difficult even as its volume falls. A staffing decision based only on transaction counts could misread the specialist capacity still required, and in aerospace that capacity is often a handful of experienced engineers and quality staff who are already scarce.
There is also a difference between capacity released and cash saved. Time returned to a sales engineer may support additional quotations or better customer discussions without reducing payroll, and earlier completion of a transaction affects billing only if acceptance and invoicing conditions also permit it. Such outcomes can be valuable, but each follows a different causal path, and the business case becomes stronger when it names the path it intends to improve and measures the resulting impact directly.
The strongest case for keeping people involved
A thoughtful objection is that extensive delegation may add little value in a low-volume, high-consequence process, where a better evidence pack for a human decision is the preferred outcome. Some transactions are also better served by conventional rules-based automation: when the inputs and the authorised response are stable, an agent's interpretive flexibility can add cost and uncertainty. The relevant comparison is therefore the best available process, including ordinary software and well-supported human work.
Deloitte's August 2026 research, based on 501 senior US leaders at organisations already at least piloting agentic AI, surveyed between April and June 2026, together with 20 executive interviews, found that many organisations are adding agents on top of existing processes because this produces the short payback periods needed to justify the investment, while fuller process redesign takes years. Only 16% of respondents described their business processes as prepared for agents. The evidence supports a measured sequence, though it cannot establish the return at a European aerospace or defence supplier (Widener et al., 2026).
The more sceptical view deserves its full weight. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, and noted that many use cases presented as agentic do not need an agentic implementation at all (Gartner, 2025). That forecast is now more than a year old and remains a prediction rather than an observed outcome, but it is a fair warning against judging a trial by how capable the agent looks in a demonstration.
Aviation regulation provides an additional reason to keep use cases distinct. In June 2026, EASA released for consultation Proposed Issue 03 of its AI concept paper, the final concept-paper deliverable under its AI Roadmap 2.0, which extends its human-centric trustworthiness guidance to further AI techniques and to advanced automation in which the human end user may be remote or absent; comments closed on 12 August 2026 (European Union Aviation Safety Agency, 2026). That guidance concerns AI in aviation safety. A procurement assistant and an AI application that affects airworthiness do not acquire the same assurance argument merely because both use AI.
The European Commission's guidance on the AI Act likewise separates obligations by category. Following the AI Omnibus, which entered into force on 27 July 2026, the rules for high-risk uses listed in Annex III apply from 2 December 2027, and the rules for high-risk AI embedded in regulated products apply from 2 August 2028. The Act requires appropriate human oversight for high-risk systems; it does not require that every agent action in an ordinary commercial workflow be approved by a person (European Commission, 2026). Legal compliance, contractual authority and internal approval policy answer different questions, and an enterprise gains clarity by naming which boundary each control protects.
What credible deployments may look like by October 2027?
My reading is that between now and October 2027, credible industrial deployments in aerospace and defence will increasingly combine agent interpretation, conventional rules-based enforcement and targeted human approval, with broad autonomy confined to narrow, low-consequence tasks. I hold that view with moderate confidence: it is more likely than not, and it is not a claim that the market will settle on a single architecture. SAP's planner-approval design, AAR's staged move towards autonomous ordering and NIST's authorisation work all point in this direction, though none of them demonstrates adoption or financial return in the sector (Metzger, 2026; AAR Corp., 2026; NIST NCCoE, 2026).
Three signals would confirm the view: published customer cases that report whole-process cycle time, exception costs and the conditions under which actions were authorised; procurement requirements from primes and defence customers that ask suppliers for scoped agent permissions and practical revocation; and whether AAR's planned negotiation agents reach customers with explicit authority limits over the coming months. Documented deployments with broad autonomy, little intervention, acceptable incident rates, and fully disclosed costs would weaken it and change the advice that follows.
The earliest gains are likely to go to suppliers with documented processes, a written delegation of authority and reliable transaction data, because those foundations shorten the work of defining and testing what an agent may do. That is an inference, not a measured ranking of companies or countries. Smaller suppliers may benefit most from vendor-managed capabilities, yet they still have to decide what the software may commit on their behalf. Organisations that measure pilots only by preparation speed are likely to see less of the benefit reach the customer. The boundary between prime and supplier matters most, because faster preparation within one company cannot, on its own, accelerate a customer's acceptance decision.
A trial that gives management a usable investment decision
Were I part of the investment discussion at a European aerospace or defence supplier considering agents in commercial or administrative work, the option I would put forward is short enough to minute: approve one bounded trial in a single repeatable workflow with an observable economic consequence, such as routine quotation release, supplier-document packs or purchase-requisition preparation, and hold any wider rollout until the trial reports. Safety-critical decisions, new supplier approvals, configuration changes, and contractual variations would remain outside its scope unless their owners explicitly designed and validated their inclusion. The trial would also give any broader AI budget a single evidence base to build on.
The first month would establish the baseline. The process owner and a finance colleague would record the actual transaction mix, preparation effort, waiting time, corrections and completion conditions, while IT, quality and export-control colleagues mapped the data and actions the agent could touch. That work has value before any agent runs, because it separates delays caused by preparation from those caused by missing information, approval capacity or customer dependencies. In the second month, the agent would work in observation or preparation mode on an agreed set of cases, with reviewers comparing its evidence packs against the existing process and recording both omissions and useful contributions. Only in the third month, and only where the evidence justified it, would it receive a narrow, time-limited permission to execute, enforced outside the model's own interpretation, with a named business owner deciding whether the trial remained within its mandate. The resources are mostly reviewer time and integration support rather than software expenditure.
The report at ninety days would show cost per correctly completed transaction, elapsed time, exception effort and correction rates against the baseline, together with the cases excluded and whether the benefit came from the agent, cleaner data or a changed approval path. I would not put a percentage saving into the business case in advance, because the public evidence does not support one for this sector. The improvement worth looking for shows up in indicators a CFO already follows: quotation turnaround and win rate, days from requisition to order, working capital tied up in parts awaiting documentation, and on-time delivery to the prime. A combined improvement can still be commercially sound, and knowing its sources tells management whether the same investment would work elsewhere.
At six months, extending the trial to a second workflow would depend on repeatable results across both ordinary and unusual cases and on a credible way to maintain controls as the software changes. By twelve months, the discussion could move from a local productivity gain to customer response, throughput or working capital, where the evidence supports it. The options stay open throughout: keep assisted preparation, broaden delegation, or use conventional automation where it yields better results.
The timing argument is practical. Demand on aerospace and defence supply chains is expected to stay high through 2027, aftermarket platforms are already moving towards autonomous ordering, and customers are beginning to ask how suppliers govern automated actions. Waiting mainly postpones learning, because the supplier later finds out which of its boundaries are genuine constraints and which simply need to be written down more clearly. Moving too broadly carries the opposite cost, with integration and oversight commitments made before the transaction economics are understood. If the baseline showed that most elapsed time is spent with the customer or in approval capacity rather than in preparation, I would advise allocating funds to approval design and conventional workflow tools first. The planner in SAP's example still has a decision to make, and the worthwhile investment is the one that helps the organisation reach that decision sooner, with better evidence and an authorised path to completion.
Questions for the C-Suite and the Board
Which workflow has enough repeatable transactions to test a measurable economic improvement within ninety days? For a business-unit leader or COO, the answer shows where evidence can accumulate without a company-wide redesign.
Which actions could be delegated within our existing commercial, financial, quality and export-control limits? Separating preparation from execution gives the process owner a concrete scope to agree with each control function.
How much elapsed time is spent in preparation, and how much in waiting, approvals, or missing information? The answer tells the COO whether an agent addresses the main constraint or improves a useful but smaller stage.
What would an exception cost in specialist time, and who would resolve it? This makes visible the engineering, quality, and commercial capacity that unusual cases require before routine volumes increase.
Would the benefit appear as cash, extra capacity, faster customer response or earlier billing? The CFO and the process owner can then measure the intended result rather than treating every hour saved as a payroll reduction.
What evidence would justify wider delegation, and what would lead us to narrow it again? Explicit expansion and withdrawal criteria give the board a reversible investment decision.
References
AAR Corp. (2026, April 21). AAR launches Airvoyant℠, an AI-driven procurement platform for airlines and MROs [Press release]. https://www.aarcorp.com/en/newsroom/press-releases/2026/aar-launches-airvoyant-sm--an-ai-driven-procurement-platform-for-airlines-and--mros/
Bayiates, A. (2026, April 24). Business and IT leaders report AI agents are scaling faster than their guardrails. Deloitte Insights. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html
Berckman, L., Shepley, S., Chavali, A., Hardin, K., Welch, S., & Dronamraju, T. (2025, November 13). 2026 aerospace and defence industry outlook. Deloitte Insights. https://www.deloitte.com/us/en/insights/industry/aerospace-defense/aerospace-and-defense-industry-outlook.html
European Commission. (2026, August 3). AI Act. Shaping Europe's Digital Future. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
European Union Aviation Safety Agency. (2026, June 3). EASA releases latest issue of its Concept Paper on Artificial Intelligence for comment. https://www.easa.europa.eu/en/newsroom-and-events/news/easa-releases-latest-issue-its-concept-paper-artificial-intelligence
Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027 [Press release]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
KPMG LLP. (2026, September 24). AI's value story sharpens as organisations gain confidence in governance, accountability and workforce adoption [Press release]. https://kpmg.com/us/en/media/news/q3-ai-pulse-2026.html
Metzger, D. (2026, April 20). SAP at Hannover Messe 2026: Operationalising agentic AI to drive resilient, end-to-end manufacturing. SAP News Centre. https://news.sap.com/2026/04/sap-at-hannover-messe-2026-agentic-ai-resilient-manufacturing/
Microsoft. (2026, May 5). Agents, human agency, and the opportunity for every organisation: 2026 Work Trend Index annual report. WorkLab. https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
National Institute of Standards and Technology, National Cybersecurity Centre of Excellence. (2026, September 29). Comments on software and agentic SI identity concept paper. https://www.nccoe.nist.gov/news-insights/comments-software-and-agentic-si-identity-concept-paper
Widener, C., Shact, L., Jarvis, D., & Mazumder, S. (2026, August 12). AI agents are only the beginning: The path to agentic transformation. Deloitte Insights. https://www.deloitte.com/us/en/insights/industry/technology/path-to-agentic-transformation.html
AI tools were used for research, analysis and editing. The analysis and final text were reviewed and approved by Andy Demir.
About Andy Demir
Andy Demir is the editor of SwissTechnic and writes on strategy, international growth and capital in aerospace, defence, space and advanced technology. Drawing on a career in commercial leadership, the analysis examines how technology, sovereignty and industrial capability shape competitive power and long-term value.
